[tunix] Add target-aligned diffusion SFT adapter

[ethannnnnn]

Motivation

Tunix now has model-agnostic diffusion batch and logits contracts, but
PeftTrainer still needs a small adapter and a target-aligned supervised loss
before those contracts can be used for diffusion CFT or SFT. Autoregressive
loss helpers cannot be reused because they shift targets by one token.

Scope

This PR adds only the diffusion SFT integration layer:

• adapt an external raw batch into DiffusionTokenBatch;
• call a model-specific DiffusionLogitsFn through the public diffusion
  contract;
• compute float32 cross entropy at the same token positions as target_ids;
• preserve the unreduced weighted numerator and denominator expected by
  PeftTrainer; and
• sanitize inactive positions so zero-weight NaN, infinity, or invalid target
  values cannot contaminate the loss or gradient.

It does not add a model implementation, tokenizer, dataset recipe,
distillation objective, rollout path, or RL trainer.

Dependencies

This PR depends on the contracts merged in #1832 and the fractional and
zero-denominator gradient behavior merged in #1833. Both dependencies are
verified against main by the guarded publisher before the patch is replayed.
It does not depend on #1850 or the optional OPD track.

Compatibility

The existing autoregressive PeftTrainer path is unchanged. Users opt in by
calling configure_diffusion_sft with their batch adapter and target-aligned
logits function. CFT and SFT use the same objective; the external adapter
chooses which target positions and weights are active.

Tests

python -m pytest -q tests/diffusion/types_test.py tests/sft/diffusion_sft_test.py
python -m pytest -q tests/sft/peft_trainer_test.py

The focused suite covers trainer wiring, same-position alignment, float32
metrics, zero-weight numerical safety and zero gradients, jitted gradients,
microbatch equivalence, and shape rejection. The guarded publisher also runs
full PeftTrainer regression tests, Pyink, scoped error/fatal Pylint, Python
compilation, and git diff --check.

Design document:
https://docs.google.com/document/d/1Xe-98ScS2RSH29AdhTdc9tO4WHQG5wCbFkJA3gmIUZQ/edit