Add end-to-end distributed inference runtime

[manishklach]

What changed

• adds a versioned, checksummed synthetic checkpoint format and deterministic expert sharding
• implements causal attention, routing, stationary INT4 SwiGLU workers, activation dispatch, and weighted reduction
• supports both in-process and framed-TCP multi-process execution with parity tests
• exposes OpenAI-compatible model and chat-completions endpoints
• adds worker/server/checkpoint CLIs, Docker Compose topology, runtime documentation, and deployment boundaries

Why

The repository previously proved individual planning, routing, quantization, and HIP-kernel pieces but did not provide a runnable request-to-response path. This vertical slice makes the architecture executable on CPU while preserving the interfaces intended for ROCm kernels and RoCE/GPUDirect transport.

Validation

• python -m pytest -q — 13 passed
• python -m ruff check . — passed
• python -m build — wheel and source distribution built
• local and TCP clusters produce matching logits in integration tests