Fix distributed CMNRL gradient synchronization

[capemox]

Summary

• move CMNRL GradCache orchestration outside the DDP-wrapped model forward, following the Sentence Transformers trainer boundary
• replay each mini-batch through the wrapped model under no_sync() and synchronize accumulated gradients on the final replay
• reject unsafe direct distributed model calls and incompatible static-graph DDP explicitly
• replace the finite-gradient smoke assertion with exact two-rank MNRL/CMNRL gradient equivalence, plus an end-to-end distributed Trainer step with gradient accumulation
• document the distributed CMNRL behavior and include it in the 0.1.1 changelog

Release note

CMNRL remains directly callable in a single process. Starting in 0.1.1, directly wrapping CachedMNRLForPretraining in DDP raises an error; distributed CMNRL must use PretenseTrainer so every cached replay is synchronized correctly.

Validation

• uv run ruff check .
• uv run mypy --python-version 3.10 src/pretense
• uv run pytest -q
• two-process distributed matrix on the current lockfile: gather on/off, find-unused on/off, direct-call guard, static-graph guard, and Trainer gradient accumulation
• full dependency-floor pytest suite and the same two-process distributed matrix on PyTorch 2.2 / the lowest supported dependencies
• wheel and sdist build plus Twine metadata checks