test(megatron): cover compiled MXFP4 backward

[gphuang]

Summary

Closes the post-merge review gap from PR #986.

Adds regression coverage for FP8/BF16 forward with MXFP4 backward through torch.compile, including exact eager/compiled output and gradient comparisons. It also covers YAML-to-FluxConfig precision propagation and runtime mode-census emission.

Test plan

• 35 config and audit tests passed
• Compiled-backward checks passed for FP8 and BF16 forward modes
• Flux 535M forward-only integration test passed
• black formatting verified
• GPU memory returned to idle

[gphuang]

Verdict: compiled-backward coverage is ready to merge

No blocking findings in the full base-to-head diff.

• Exact eager/compiled outputs and input/weight gradients passed for FP8 and BF16 forward with MXFP4 backward.
• Config propagation, runtime audit emission, and the Flux integration path passed on GPU.
• Required CI is green; GPU CI is skipped, with the equivalent targeted GPU gate run on the allocated node.

Two optional hardening ideas—testing a future non-default backward mode and pinning preshuffle state inside the test—do not affect the current production path.