fix(train): reconstruct MixedDataLoader RNG stream on resume

[AuraIis]

Root cause

MixedDataLoader owns independent NumPy generators for each language plus a shuffle generator and fractional row-credit state. Checkpoints restore global RNGs but not these loader-owned streams, so resumed training restarted sampling at batch zero.

Measured on the pinned clean checkout: LR and eval loss restored, while train loss diverged immediately; an isolated token-stream comparison confirmed different post-resume batches.

Fix

• count mixed batches consumed by the loader
• add a data-free deterministic fast-forward that reconstructs:
  • each per-language RNG
  • the batch-row shuffle RNG
  • fractional mix row credit
• after loading a checkpoint, advance by
  trainer_state.step * gradient_accumulation
• reconstruction uses each rank's own seed/shard, so DDP ranks recover their own streams
• no token data is read during fast-forward

This works with the existing step_60000 checkpoint format; no migration or checkpoint rewrite is required.

Regression coverage

• exact next-batch equality after 160 consumed batches (step 20 × grad accumulation 8)
• ten subsequent batches remain byte-identical
• rank-specific DDP-shard reconstruction
• invalid negative fast-forward rejected

The existing code/config/tokenizer provenance checks remain in force; changing loader configuration across a resume is still a provenance mismatch.