chore(m11): offload-audit GPU memory probes for vast smoke runs

[zhenyulincs]

What

Probe-only instrumentation from the M11 offload audit (4x RTX 5090, cu130 vast instance) — the diagnostics used to root-cause the ~12 GB whole-GPU residual (torch-hook tms pauses only the init-time 2.76 GB; optimizer state never offloads) that led to rlops/rlix#17 and rlops/rlix#31.

• memory_utils: available_memory() now also reports the physical GPU id (CVD slice lookup) so per-GPU logs are unambiguous when several actors each see their GPU as cuda:0; new log_nontorch(label) probe logs whole_used / torch reserved / non_torch so deltas between consecutive probes attribute non-torch allocations (NCCL, context, driver pools) phase by phase.
• megatron actor: log_nontorch around reload_process_groups (wake path) and build_cpu_bucket_cache.
• megatron model: log_nontorch around forward_backward and optimizer.step in train_one_step.
• sglang_engine: _log_whole_gpu(label) — whole-GPU used via nvidia-smi for the engine's visible GPUs, before/after release_memory_occupation / resume_memory_occupation. Deliberately whole-GPU (per-process attribution breaks under some container PID namespaces) and deliberately print (engine actor processes have no logging handler; Ray forwards stdout unconditionally).
• docker/patch/latest/sglang.patch: strip a trailing space on one context line — without it the patch fails to apply on the vast cu130 image.

What this is NOT

• No behavior change beyond added logging (verified: the only non-logging diff is the sglang.patch whitespace byte).
• The run_sync_session supersession guard observed during the same audit is not included here — that fix is fix(miles): supersession guard in run_sync_session — sync latest when requested step is stale #34 (approved separately).

Validation

• All probes exercised on the vast 4x5090 dual-pipeline smoke (EXIT_CODE=0); probe output is the evidence base for the residual-threshold work in feat(rlix): lower whole-GPU residual gate default to 7.0 GiB; preload dual smoke rlix#31.
• python -m py_compile on all four touched Python files.

[zhenyulincs]

Closing — wrong content: this PR carried the audit instrumentation probes, but the intended contribution is the vast-compute sync/debug script toolkit, which will be PR'd to rlops/rlix instead. The probe diff remains available locally (stash) if ever needed.