Cosmos3-edge PR 3: /generate over a persistent WebSocket for control loops

[merceod]

Adds /generate/ws, the native /generate request over one persistent WebSocket, for clients that run a control loop against the server (robot policies, streaming world models) and can't afford an HTTP round trip per step.

• One message = one request with the /generate fields (text, files as [{"name","data"}], input_modalities, output_modalities, model_kwargs, request_id). JSON text frames carry data as base64, msgpack binary frames carry raw bytes, so images and actions skip base64.
• Replies use the message's encoding, one frame per ResultChunk ({"request_id","modality","data","metadata"}) followed by {"request_id","finish": true}. Errors travel in-band as {"request_id","error"} and the socket stays open.
• Messages pipeline. Each one gets its own task, replies interleave and carry the request_id. Closing the socket cancels the tasks, and tearing down iter_result_chunks aborts the engine request the same way /generate streaming does.
• Media are persisted under upload_dir with the same layout as the multipart route (shared _decode_ws_files / _ws_input_layout), so models see identical file_paths / prompt_parts. submit_request and iter_result_chunks are reused unchanged.

Also adds the websockets dependency (uvicorn needs a WS implementation), a section in docs/clients.rst, and examples/cosmos3_action_ws_client.py, an openpi-style policy client that streams observations, decodes float32 [chunk, action_dim] replies and reports chunks/s, actions/s and latency percentiles.

Tests in test/modular/test_generate_ws.py (Starlette in-process client against a fake API server) cover a JSON round trip with media persisted in order, three pipelined msgpack requests, rejected messages answered in-band without dropping the socket, the not-ready close (1013), undecodable / non-object / empty frames, and a mid-stream disconnect cancelling the request. ruff and test/modular pass.