feat: assemble motion tracking Numba plan

[TATP-233]

Fixes #923

Summary

• register motion tracking preamble, reward, observation, and termination Numba item functions on the existing resolved term plan
• extend the shared fused assembler narrowly for bool/int scalar parameters, typed tuple parameters, parameter-schema cache keys, and cold-path workspace view binding
• replace the motion tracking hand-written full aggregator and fixed term order with the automatic fused plan while preserving task-owned noise and lifecycle orchestration
• reuse preallocated actor/critic, reward, termination, and shared workspace buffers; remove the legacy accelerator path
• update the focused benchmark from reward/termination wording to full fused plan-state measurements

Owner boundaries

• no backend, runner, NpEnv lifecycle, training ABI, checkpoint, sim2sim, asset, or regular CI changes
• motion reference sampling, adaptive failure accounting, frame advance, clip-end behavior, teleport, and final observation remain task-owned
• this PR completes only the G1 motion-tracking pilot; it does not promote other task families

Validation

• make test-all
  • 1667 passed, 37 skipped, 280 deselected, 1 xfailed
  • Ruff, mypy, Pyright, and benchmark import smoke passed
• uv run pytest -q tests/envs/test_g1_motion_tracking_numba.py -m slow
  • 2 passed
• uv run pytest -q tests/envs/test_env_configs.py -k 'g1_motion_tracking_clip_end'
  • 5 passed
• shared assembler execution now explicitly covers bool/int scalar and typed tuple parameters
• actor owner-noise / clean critic boundary, output buffer reuse, and first-JIT reporting are covered

Performance

Command:

uv run python -m benchmark.env.benchmark_g1_motion_tracking_numba --profiles sac_default --num-envs 2048 8192 --threads 8 --iters 80 --warmup 8 --no-e2e

8-thread full fused plan-state results versus the pre-migration hand-written _compute_update_state_kernel baseline:

envs	current	old baseline	retained
2048	10.43M env/s	10.72M env/s	97.3%
8192	12.24M env/s	11.60M env/s	105.5%

Both exceed the 90% acceptance gate. Parity maxima were about 1.0e-7 for reward and 7.2e-7 for critic observation, with zero termination mismatches.

Cold/cache measurements at N=2048, 8 threads:

• cache miss assembly: 0.813 ms; first JIT execute: 11876.861 ms
• cache-hit assembly: 0.001 ms; first warm execute: 0.198 ms

Scope size

• 9 files
• +640 / -1045 lines