fix(train): replay epochs after prevalidation rollback

[onion-hong]

When an epoch-end callback makes the online model non-finite, _recover_before_validation() restores the healthy checkpoint but returns false. Validation then succeeds on restored weights, so the rolled-back attempt is logged as a completed epoch, the recovery counter is cleared, and the MoE scheduler receives recovered=False. A repeated fault can therefore silently discard each epoch without exhausting the retry limit.

Return the replay signal after an online rollback, reset gradient accumulation before replay, and finalize that attempt with recovered=True, validated=False. EMA-only resynchronization keeps its existing behavior. Adjust the existing online/EMA test to inject once; add offline real-training regression cases for one-time and persistent faults, checking optimizer updates, validation calls, CSV rows, scheduler flags and the recovery budget.

Validation:

• New regression cases fail on the original code (no replay; persistent faults do not raise).
• On isolated upstream 3bd0a60 plus this patch: recovery/NaN/MoE scheduler tests: 72 passed, 1 skipped.
• Related engine EMA-resync and online/EMA recovery tests: 2 passed.
• Ruff format and git diff --check pass; modified trainer retains 11 pre-existing lint diagnostics.

This fixes prevalidation rollback accounting; it does not prevent the original source of non-finite values.

[isLinXu]

LGTM