fix(train): resume at exact global step instead of flooring to grad-accum boundary

[longzhenren]

Problem

compute_resume_position floors skip to the previous grad-accum boundary when
global_step % grad_accum != 0, and pulls aligned_global_step back to that
boundary. With grad_accum=4, batches_per_epoch=10, resuming at global_step=12 it
returns (epoch=1, skip=0, aligned=10) instead of (1, 2, 12).

The premise for the floor is that a resumed run must start a fresh accumulation
cycle. That premise does not hold: Accelerate's load_state restores the
accumulation counter, so resume already continues from the exact checkpointed
position.

Root cause

Verified empirically (CPU, accelerate 1.12): after 3 micro-batches under
accelerator.accumulate, save_state/load_state round-trips
accelerator.step (3 -> 3). save_state writes self.step and load_state
assigns it back. The training loop resumes through
accelerator.load_state in load_full_state.

With the floor, epoch-1 batches 0-1 are re-fed after resume, their gradients are
counted into the next optimizer step, and the per-step seed (keyed on aligned
global_step) shifts. The same floor also corrupts the derived opt_step
(10//4=2 vs the checkpoint's real 12//4=3), shifting the LR schedule.

Fix

• skip = global_step % batches_per_epoch, aligned_global_step = global_step
  (drop the floor). grad_accum stays in the signature but is a no-op.
• grad_accum=1 behavior is unchanged.
• Tests updated to assert the exact resume position, plus a new mid-accumulation
  case (ga=4, bpe=10, G=12 -> skip=2, aligned=12).

Validation

• pytest tests/test_checkpointing.py: 23 passed.
• Train-adjacent subset (checkpointing, optimizer groups, training utils,
  prepare/unwrap, seeding, dataloader seed): 59 passed.
• ruff check: clean.