feat(dr): generic interval term dispatch in DomainRandomizationManager

[TATP-233]

Summary

UniLab half of #1502 (approved approach A: descriptor/operation registry). Consumes the unisim-core contract PR unilabsim/unisim#21.

• DomainRandomizationManager.apply_interval_randomization_if_due is now fully generic (src/unilab/dr/manager.py:91): no interval term names, no per-field capability if branches. It validates plan.iter_ops() against capabilities.get_unsupported_interval_terms(...) and fails closed with backend type + sorted term names, then dispatches once.
• Builtin providers migrated to descriptors: locomotion push (src/unilab/dr/dr_utils.py) and Sharpa body force (src/unilab/tasks/manipulation/sharpa_inhand/rotation.py) now emit IntervalTermOp entries with identical payloads/timing; validation uses supports_interval_term.
• Legacy IntervalRandomizationPlan fields keep working through iter_ops() (deprecated in the provider docstring; removal in the next unisim-core major) — existing task owners and the Entity/EventManager interval path (src/unilab/base/entity.py, src/unilab/envs/mdp/events.py) are untouched and run through the compatibility adapter.
• unilab.dr re-exports IntervalTermOp + INTERVAL_TERM_*.
• Docs updated in en + zh_CN: DR contract (term registration, capability ownership, pickle/spawn rules, fail-closed) and provider authoring guide with migration note.
• pyproject.toml: unisim-core>=0.1.14 → >=1.1.0.

Tests

• New interval-side coverage in tests/dr/test_manager.py: custom-term extensibility without touching the manager (acceptance proof), capability-negative errors (term + backend in message), legacy-fields compat, mixed legacy+ops, empty/None plans, multi-op single dispatch, pickle protocol-4 round-trip.
• Intended behavior change from unisim-core 1.1.0: mjwarp/genesis previously silently dropped body_torque / body_angular_velocity_delta; they now fail closed. No UniLab test relied on the silent behavior; three backend tests were updated for the new unified error messages/validation order (test_genesis_backend.py, test_isaacgym_backend.py, test_sim_backend_smoke.py).

Benchmark (interval hot path, 20k calls, min of 5 repeats, this machine)

Path	Baseline (main + unisim-core 0.1.14)	This PR (+ unisim-core 1.1.0)
Manager dispatch (fake backend, push+force plan)	884 ns/call	1625 ns/call
MuJoCo apply_interval_randomization (body-force plan)	1947 ns/call	2200 ns/call

The generic dispatch adds ≤0.9 µs per interval application; interval terms fire on multi-step cadence (e.g. push interval) against ~ms physics steps, so there is no measurable hot-path regression and no extra backend round-trip (still one apply_interval_randomization call per due step). An ops-only fast path in iter_ops() avoids per-call tuple re-allocation (unisim b49e8b5).

Validation

• UV_NO_SYNC=1 make test-all: ruff format/check clean, mypy clean (148 files), pyright 0 errors, pytest not slow 1578 passed / 20 skipped / 1 xfailed, benchmark smoke 36/37 pass (1 platform-optional mlx skip).
• One expected local failure: tests/base/test_backend_imports.py::test_unisim_dependency_is_installed_from_package_index — by design it rejects non-index unisim-core installs; local verification used an editable install of feat(dr): declarative interval term descriptors with generic backend dispatch unilabsim/unisim#21. It passes once unisim-core 1.1.0 is published and installed from PyPI.

Dependency / merge order

1. Merge feat(dr): declarative interval term descriptors with generic backend dispatch unilabsim/unisim#21, maintainer publishes unisim-core 1.1.0 (git tag v1.1.0, trusted publishing per unisim docs/release.md).
2. Then regenerate uv.lock here (uv lock) — deliberately not done yet since 1.1.0 is not on PyPI; remote CI on this PR will be red until then. I will push the lock update and confirm CI green after the release.