Fix Matrix-Game-3 Ulysses attention dispatch

[zjn20030811]

Summary

• Fix the Matrix-Game-3 Ulysses adapter to pass the attention backend selector
  using the keyword accepted by wan.modules.attention.attention.
• Add a CPU regression that exercises the adapter with a strict attention stub,
  so a renamed or misspelled keyword fails before multi-GPU inference.

Fixes #80

Why

The distributed path passed fa_version=... to a function whose parameter is
version. Every run with --ulysses_size > 1 therefore raised TypeError in
the first attention block, before any backend selection or model computation.
The single-word forwarding fix leaves attention math and all non-Ulysses paths
unchanged.

Validation

python -m pytest -q Matrix-Game-3/tests/test_ulysses_attention.py
1 passed
python -m py_compile Matrix-Game-3/wan/distributed/ulysses.py Matrix-Game-3/tests/test_ulysses_attention.py
git diff --check