Fix EMA updates across gradient accumulation and resume

[zjn20030811]

Summary

• update the EMA model only after a synchronized optimizer step that was not skipped for overflow
• persist the EMA optimization counter next to its averaged weights and restore it on resume
• retain compatibility with older checkpoints by falling back to the checkpoint global_step when the sidecar is absent

Motivation

The training loop currently calls ema_model.step() after every microbatch. Accelerate suppresses optimizer.step() inside a gradient-accumulation window, but the standalone EMA update is not suppressed. With accumulation factor N, this advances the EMA warmup schedule N times per optimizer update, repeatedly averages unchanged parameters, and performs N full traversals of the 1B-parameter model.

Resume has a separate schedule discontinuity: only the averaged weights are loaded, while EMAModel.optimization_step starts again at zero. The next EMA update therefore uses zero decay and overwrites the loaded history with the online model.

This change keeps the existing global_step data-consumption semantics while tracking successful EMA updates independently. New checkpoints resume the exact schedule; old checkpoints continue to load with an explicit best-effort migration.

Validation

• python -m pytest tests/test_ema_model.py -q — 6 passed with Accelerate 0.30.1
• real CPU accumulation test confirms updates occur only on microbatches 4 and 8 for accumulation factor 4
• overflow-skip regression leaves both EMA weights and counter unchanged
• save/load continuation matches an uninterrupted EMA trajectory exactly
• full Ruff check for tests, fatal-rule checks for changed source files, py_compile, and git diff --check