perf: enable multi-GPU GPU-resident replay for SAC

[TATP-233]

Summary

Enable rank-local GPU-resident replay for multi-GPU SAC off-policy training on AMD ROCm. Each learner rank owns an independent GPU-resident mirror of the shared CPU replay buffer, eliminating per-tick collector CPU pack and per-rank batch H2D.

Fixes #813

Changes

• New pipeline: MultiGPUGPUResidentReplayPipeline wraps the single-GPU GPUResidentReplayPipeline (feat(replay): opt-in GPU-resident replay pipeline for single-GPU off-policy #806) per rank, providing the same interface as MultiGPUCPUPinnedReplayPipeline without collector pack IPC.
• Runner wiring: MultiGPUOffPolicyRunner selects pipeline via training.replay_pipeline; GPU-resident path skips collector pack queues/slots.
• Config unlock: remove the single-GPU-only restriction in scripts/train_offpolicy.py for training.replay_pipeline=gpu_resident.
• ROCm test compatibility: update repo hygiene / torch-source tests to accept both CUDA and ROCm configs so make test-all passes with pyproject.rocm.toml active.
• Benchmark helper: add benchmark/rl/extract_offpolicy_metrics.py for TensorBoard metric extraction.

Benchmarks (8x AMD Instinct MI300X, ROCm 7.2)

Task: g1_motion_tracking / motrix, 30 iterations, last-20 average.

GPUs	Pipeline	iter_ms	steps/s	Collector/s	collector replay_ms	collector wait_ms	rank barrier_ms
2	cpu_pinned	77.0	27,420	28,982	2.00	2.75	2.69
2	gpu_resident	81.5	25,482	24,751	1.67	4.96	2.38
4	cpu_pinned	86.4	24,815	22,556	1.95	11.19	2.19
4	gpu_resident	75.8	27,544	27,777	1.79	2.36	2.23
8	cpu_pinned	118.4	19,536	18,477	1.83	44.66	1.49
8	gpu_resident	86.2	25,146	24,802	1.75	12.71	2.31

Key results:

• collector replay_ms stays 1.7–1.8 ms across 2/4/8 GPU (no rank degradation, DoD met).
• 8 GPU vs CPU-pinned: iter_ms -27% (118.4 → 86.2), steps/s +29% (19,536 → 25,146), learner collector wait -72% (44.7 → 12.7 ms).

Microbenchmark baseline (host replay harness, capacity 1M rows): CPU sample+H2D 7.0 → 10.8 → 20.6 ms (1→4→8 GPU); GPU-resident sample 0.14 → 0.50 → 0.66 ms.

Validation

• make test-all passed: 1634 passed, 26 skipped, 267 deselected, 1 xfailed.
• Benchmark smoke test passed (34/35 module-mode, 35/36 script-mode; only mlx optional skipped).
• Relevant unit tests: tests/ipc/test_replay_pipeline_gpu_resident.py, tests/ipc/test_multi_gpu_replay_pack.py, tests/algos/test_offpolicy_double_buffer_runner.py, tests/algos/test_offpolicy_runner_unit.py all pass.
• End-to-end smoke: 2/4/8 GPU g1_motion_tracking / motrix training completes successfully.

Notes

• Sharded replay view, periodic bulk refresh, field-slim, and freshness relaxation are left as follow-up work per perf: SAC replay 多卡（2/4/8 GPU）traffic scaling 与 rank-local GPU cache 验证 — 自 #694 拆出 #813 scope.
• Local pyproject.toml / uv.lock were swapped by make sync-rocm for AMD development and are intentionally not part of this PR.