feat: add mix bb.from_urdf URDF importer

[jimsynz]

Closes #101.

Summary

Adds a mix bb.from_urdf Igniter task that reads a URDF XML file and generates an equivalent BB DSL module:

mix bb.from_urdf path/to/robot.urdf --module MyApp.Robot

Three pieces:

• BB.Urdf.Parser — parses URDF XML into a plain-map representation via :xmerl_scan (no extra deps). Collects warnings for <mimic>, <safety_controller>, <transmission>, and <gazebo> blocks rather than failing the import, so partial fidelity is preferred over rejecting real-world URDFs.
• BB.Urdf.Importer — converts parsed URDF into a quoted BB DSL module and formats it via Sourceror with ~u sigil literals for unit-bearing values. Joints are nested under their parent links in topology order; the root link is identified as the one never appearing as a <child>. Cardinal axis vectors map to BB's axis do roll/pitch end Euler form (0 1 0 → roll ~u(-90 degree), etc.); arbitrary vectors get rotated to Z via computed Euler angles.
• mix bb.from_urdf — Igniter task that wraps the two, writes the module via Igniter.Project.Module.create_module, and surfaces parser warnings via Igniter.add_warning/2.

What's deliberately not in this PR

• Mesh file management. Mesh filename attributes (including package:// URIs) are passed through verbatim. The user copies meshes and rewrites paths themselves; that felt like a separate concern from generation.
• <mimic> / <safety_controller> / <transmission> translation. No clean BB analogue — surfaced as warnings instead.
• Tutorial doc. 07-urdf-import.md from the issue's "Related" section. The existing 07-parameters.md makes the numbering awkward; happy to add this as a follow-up once we decide on a slot (e.g. renumber, or move to documentation/how-to/).
• Regenerating bb_example_wx200's definition from a vendor URDF as a smoke test, per the issue's open question. Round-trip works on the test fixture (URDF → DSL → compile → URDF), but adding the example regen feels like a separate PR.

Test plan

• mix test test/bb/urdf/parser_test.exs (8 tests)
• mix test test/bb/urdf/importer_test.exs (10 tests, including a generated-source-compiles check)
• mix test test/mix/tasks/bb.from_urdf_test.exs (4 Igniter tests)
• mix test (938 tests, no regressions)
• mix format --check-formatted
• mix credo --strict
• mix dialyzer (clean)
• Manual round-trip on two_link_arm.urdf: parse → generate DSL → compile → export → same structure

[jimsynz]

Tested against real-world URDFs

Ran the importer against four publicly-available URDFs to shake out edge cases:

URDF	Source	Links / Joints	Result
Universal Robots UR5	ros-industrial	9 / 8	✅ parse → generate → compile → round-trip
Franka Panda	StanfordASL mirror	12 / 11	✅ parse → generate → compile → round-trip
ANYmal B	ANYbotics simple desc	22 / 21	✅ parse → generate → compile → round-trip
Allegro Hand	simlabrobotics	21 / 20	✅ parse → generate → compile → round-trip

Round-trip = parse → generate DSL → compile → BB.Urdf.Exporter.export/1, with link/joint counts preserved.

Two real issues surfaced and are fixed in 1092ec7:

• URDF lets many visuals reference one named <material>, but BB's DSL requires globally-unique entity names — the four corpus modules all failed to compile because the generated source had things like four material name(:panda_white) blocks. Fix: keep the URDF name on the first visual that uses each material; subsequent occurrences emit anonymous materials (BB auto-generates an identifier when name is absent). Regression test added with a synthetic fixture.
• BB.Dsl.Mesh.scale defaults to the integer 1, and the URDF exporter calls :erlang.float_to_binary/2 on it, which crashes. Always emit scale as a float so the generated module round-trips. (The exporter bug is pre-existing — happy to file a separate issue if useful.)

Parser warnings observed on the corpus all look correct: <safety_controller> on every Panda joint, the <mimic> on the Panda gripper finger, <transmission> on every UR5 actuator, <gazebo> extensions on ANYmal, and Allegro's repeated material "black" referenced by name but not defined at top level for inline material references with no top-level def.

[jimsynz]

Good catch — BB.Sensor.Mimic is documented as the URDF mimic equivalent (same position * multiplier + offset formula). Wired it up in 3159a56: <mimic> joints now emit a BB.Sensor.Mimic attached to the mimicking joint instead of getting skipped with a warning.

For example, the Panda's panda_finger_joint2 now generates:

joint :panda_finger_joint2 do
  type :prismatic
  ...
  sensor :panda_finger_joint2_mimic,
         {BB.Sensor.Mimic, source: :panda_finger_joint1}

  link :panda_rightfinger
end

Multiplier and offset are omitted when they match URDF's defaults of 1.0 and 0.0. Updated tests on the existing mimic_and_transmission.urdf fixture (which has multiplier="-1") verify the non-default case.

[jimsynz]

Expanded corpus

Ran against eleven public URDFs spanning industrial arms, quadrupeds, mobile robots, and simple physics models. Two new failure modes surfaced (fixed in ecc341c) before everything compiled and round-tripped:

URDF	Source	Links / Joints	Notes
Allegro Hand	simlabrobotics	21 / 20	inline material refs without top-level def → parser warnings
ANYmal B	ANYbotics simple desc	22 / 21	<gazebo> extensions skipped
Cartpole	Bullet	3 / 2	clean
Drake acrobot	RobotLocomotion	3 / 3	<joint parent="world"> dropped via new world-anchor pass
Drake pendulum	RobotLocomotion	4 / 3	clean
KUKA iiwa	Bullet	8 / 7	clean
Bullet minitaur	Bullet	27 / 26	hit the new joint/link name-collision rename
Franka Panda	StanfordASL	12 / 11	<mimic> → BB.Sensor.Mimic, <safety_controller> skipped
SO-101	TheRobotStudio	7 / 6	name-collision rename for the gripper link/joint pair
UR5	ros-industrial	9 / 8	<transmission> and <gazebo> skipped
WidowX	Interbotix	8 / 7	xacro file; non-xacro portions form a valid robot

The two new pre-passes:

• World anchor: URDFs commonly anchor the base to a synthetic world link with a fixed joint. BB has no world-frame concept — the topology root is the robot. Such joints are now dropped with a warning, leaving the child unparented so the root-finder picks it up.
• Name collision: Several URDFs reuse a name across link and joint namespaces (so101's gripper, minitaur's motor_back_rightL_link). Colliding joints get renamed with a _joint suffix and <mimic> source references are rewritten to match.

Regression tests added with synthetic fixtures for both cases.