feat(reset): publish call-graph diagnostics

[TATP-233]

Why

#1874 has exhausted reset transport and tensor-expression micro-optimizations: #1913, #1914, #1940, UniSim #332, and #1941 were all neutral, while phase timings repeatedly reflected queued GPU drain rather than Python dispatch. Before implementing an owner-level fused reset transaction, the call graph itself must be measured rather than inferred from wall-time labels.

Change

Add stable selected-reset count diagnostics alongside existing timing keys:

• reset_done_event_term_count — configured reset-mode event terms.
• reset_done_command_term_count — configured command terms.
• reset_done_manager_reset_count — Manager reset passes invoked (currently seven).
• reset_done_observation_term_count — observation terms declared across groups.
• reset_done_sampler_host_transfer_count — explicit sampler D2H transfers in TensorMotionCommand reset:
  • one row-selector D2H for adaptive reset sampling;
  • three mirror D2H transfers during per-step frame advance (frames, clip indices, clip ends).

Implementation:

• Extend the stable reset schema in backend_timing.py.
• Publish counts from ManagerBasedRlEnv.reset().
• Add an opt-in CommandManager.reset_diagnostics() collection boundary.
• TensorMotionCommand resets its per-reset transfer counter before resampling and exposes the count through the command boundary.
• Collector benchmark treats these as count samples, aggregated like reset_done_count.
• Add a real MJWarp rollout test that forces all rows done and asserts the canonical FlashSAC motion call graph:
  • 0 reset event terms (DR-free owner);
  • 1 command term;
  • 7 manager reset passes;
  • 17 observation terms;
  • at least one sampler host transfer.

No reset semantics, manager execution, term dispatch, RNG, observations, or backend APIs change.

Validation

• UNILAB_LOCAL_UNISIM=... uv run pytest -q tests/base/test_torch_env.py tests/managers tests/envs/test_motion_command_partial_reset.py tests/envs/mdp/test_uniform_velocity_command.py — 137 passed.
• Real MJWarp call-graph rollout — passed.
• Full non-slow suite — 1815 passed / 20 skipped / 566 deselected.
• make check — passed (existing cli.py Drake import warning only).

Intended use

This is diagnostic-only and deliberately performance-neutral. The next 592 run can now report Manager dispatch counts per reset row, and the fused-reset contract can prove that a candidate removes Manager passes instead of merely shifting a timing boundary.