feat(megatron): report peak memory for log_probs and actor_train

[EazyReal]

Summary

• print_memory samples current allocation only at fixed lifecycle points, so it cannot expose a phase's high-water mark and an OOM can happen before the next report
• add accelerator-aware report_peak_memory(phase): reset peak stats on entry and log peak allocated/reserved memory in a finally block, so an OOMing phase still reports its high-water mark
• wrap each executed ref/teacher/actor log_probs forward and actor_train, composed with the existing timer scopes; a reused or skipped actor forward emits no separate log_probs report
• support the allocator APIs exposed by Slime's CUDA/ROCm and MUSA backends; registered accelerators without those APIs continue without peak reporting
• intentionally scope the reset to the current device and require non-nested scopes, because an inner reset would discard an outer phase's peak

Measured on a 128x H100 DeepSeek-V4-Flash run at 262,144-token context: the finally-path report attributed a 132k-context log_probs OOM to 25 GiB of retained activations. The OOM traceback alone did not show the phase's high-water mark.

Testing

• python tests/test_accelerator.py — 12 passed; covers reset-before-body, exact GB/log formatting, reporting on exceptions without swallowing them, unsupported-accelerator fallback, and CUDA/MUSA capability detection
• black --line-length 119, ruff check, and isort --profile=black pass on all touched files