Fix handling of lm head for Qwen3.5

Add FlashHeadQwen3_5 support

Simplify loading, remove duplicate indices calc, fix prefill path

- Remove clustering_config.json validation from _get_centroids (rely on
  safetensors contents directly)
- Auto-detect n_clusters from centroids tensor shape instead of requiring
  it as a parameter
- Infer vocab_size/hidden_size from weight shape instead of config metadata
- Return indices from _get_cluster_logits to avoid recomputing them in
  get_next_token (removes duplicate index_select + flatten + unique)
- Fix prefill regression: only use FlashHead for single-token decode
  (shape[0] == 1); let vLLM handle prefill natively via compiled path
- Fix sampling softmax to slice [:, -1, :] before temperature scaling

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Remove duplicate indices calc

Bump version to 0.1.5

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Bump version to 0.1.4

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Bump version to 0.1.3, add HF collection link, remove Homepage URL

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Bump version to 0.1.2, fix README images for PyPI

Use absolute URLs for images so they render on PyPI.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Bump version to 0.1.1, add PyPI metadata and landing page

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

Fix pypi-publish permissions: add contents read for release download

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>