Diffuse v0.2.9

Models that see, listen, speak, and models that answer by denoising, all
run split across machines now. The wire got much smaller, nodes hand
activations to each other instead of through the client, and there is a
prebuilt binary for Linux x86_64 and aarch64, macOS Intel and Apple
Silicon, and Windows x86_64.

Measured on Wan2.1-T2V-1.3B over two nodes: a model split across machines
is bit-identical to the same model run whole. Measured on MusicGen-small:
the waveform matches transformers' own generate to one step of 16-bit PCM.

Full notes in CHANGELOG.md.

Add a VitePress docs site with GitHub Pages deploy, and a Windows ins…

…taller

Merge pull request #11 from UnlikedOne/feat/openai-api

Feat/OpenAI api

Merge pull request #10 from UnlikedOne/fix/version-and-incomplete-models

Follow the Git tag for the version, and detect partially-held models

Check only the parameters a slice actually needs and detach the rest

Track sequence position per session instead of reading it from the cache

Give every query its own session and clear the KV cache when it ends

Load weights in their native dtype and convert activations at the tra…

…nsport boundary

Merge pull request #9 from UnlikedOne/fix/peer-identity

Fix/peer identity

README: professional rewrite, document NAT relay and streaming