chore: remove temporary MBA profiling scaffolding

[TATP-233]

Fixes #1281

Scope

Remove the temporary fine-grained MBA profiling introduced for #1255/#1256 while preserving production behavior and the retained coarse-grained timing contract.

• Removed mba_reset_*, mba_update_*, per-term/getter/observation phase timing, dynamic timing dictionaries/zero-fill, EventManager reset timing, getter recorder, and benchmark display/tests.
• Preserved physics_ms, update_state_ms, reset_done_ms, direct dr_reset_*, backend timing/behavior, the U1 state-read cache, reset transaction, R2 row-scoped reset observation, lifecycle/RNG/numerics.
• No manager/backend contract, public execution path, configuration, or sim2sim change.

The final diff is 16 files, 20 insertions, and 700 deletions. The issue's 15-file stop condition was reached by four one-line/tiny fixture cleanups required to remove dead recorder wiring; the maintainer explicitly authorized the 16-file exception before proceeding.

Backend behavior impact

No intended MuJoCo or Motrix behavior change. Exact same-seed/action parity was checked across initial observation, normal steps, partial reset observation, reward, termination/truncation, final observation, and post-reset step, including two forced timeout/autoresets per backend:

• MuJoCo: 37 arrays bit-exact, maximum absolute difference 0.0.
• Motrix: 37 arrays bit-exact, maximum absolute difference 0.0.

Validation

• make test-all
  • ruff passed
  • mypy passed
  • pyright passed
  • pytest: 2221 passed, 28 skipped, 277 deselected, 1 xfailed
  • benchmark import smoke passed
• Focused tests: 212 passed.
• Runtime/benchmark negative assertions confirm no mba_ timing key remains.
• Final commit: 65546bbe45af515262939af19b463f3f2fdf2afc.

#1281 before/after throughput

Same host and parameters, three independent serial runs per side: 8192 envs, warmup 10, measure 100, replay capacity 64, Torch threads 8. Values are median Collector/s with [min, max]; delta is after/before median.

Case	Before 5215ddcd	After 65546bbe	Delta
flashsac/go2_joystick_flat/mujoco	252,308 [250,230, 254,950]	252,144 [250,385, 255,303]	-0.07%
flashsac/g1_walk_flat/mujoco	88,465 [87,586, 88,830]	88,009 [87,466, 88,536]	-0.52%
flashsac/g1_walk_flat/motrix	85,873 [84,595, 87,145]	87,014 [83,752, 88,296]	+1.33%
sac/g1_walk_flat/mujoco	91,193 [90,560, 92,178]	91,442 [90,133, 91,527]	+0.27%
sac/g1_walk_flat/motrix	85,313 [82,638, 86,595]	83,194 [82,090, 85,255]	-2.48%
sac/g1_motion_tracking/mujoco	70,035 [68,937, 70,769]	70,120 [69,946, 70,577]	+0.12%
sac/g1_motion_tracking/motrix	63,603 [50,473, 64,000]	63,244 [63,167, 63,333]	-0.56%

The pre-cleanup sac/g1_motion_tracking/motrix series retained the known high-drift outlier (50,473); the post-cleanup series was 63,167/63,333/63,244. Raw JSON/CSV files are saved under the benchmark output directory and intentionally remain untracked.