Add an opt-in ZeRO-1/2 gradient norm fast path

[yh0903]

Summary

• add zero_optimization.compute_grad_norm, defaulting to true
• allow ZeRO Stage 1/2 GPU optimizers to skip an unused global gradient norm while preserving overflow detection
• skip the gradient multiply when clipping is disabled and the effective loss scale is exactly 1
• return None from get_global_grad_norm() when norm computation is explicitly disabled
• fail fast for gradient clipping, optimizer offload, ZenFlow, unsupported ZeRO stages, and checkpoint-restored clipping

Safety contract

Finite/overflow checking is unchanged. This does not remove the required non-finite scan; it only removes work with no mathematical consumer. The default path is unchanged.

Performance

Qwen3-30B-A3B, 48 layers, EP16 on 2x8 H100, sequence length 1024, BF16, ZeRO-1, activation checkpointing enabled:

Metric	Baseline	Fast path	Change
Median step	987.92 ms	966.67 ms	-2.2%
Optimizer update	84.93 ms	59.82 ms	-29.6%
Global norm	20.30 ms	0 ms	removed
Identity unscale	5.09 ms	0.03 ms	removed

The paired block used a benchmark-only control that restores the pre-change multiply-by-one path. Peak allocated and reserved memory were unchanged. Two additional rotated blocks isolating the opt-in norm change were both positive (+3.3% and +0.6% E2E); their optimizer-phase saving was stable at about 20 ms, while full-step variance remained larger.

Testing Done

• H100 targeted suite: 19 passed, covering Stage 1/2 update parity, overflow skip behavior, get_global_grad_norm(), clipping/offload/ZenFlow guards, checkpoint clipping, and identity/non-identity scaling
• changed-file yapf, check-torchdist, license, codespell, and other pre-commit hooks
• flake8 5.0.4 under Python 3.12; the local pre-commit Python 3.14 environment fails inside pyflakes before linting because ast.Str was removed
• canonical paired executions: fdf729ed8adac4ebea4b, feb0641f09c1d42fcb7b, ff508156ab98444ddb2f