feat: add native Kimi-K3 inference support

[earlvanze]

Closes #92.

What changed

• adds native Kimi-K3 config parsing, model registration, KDA/MLA attention,
  latent-MoE routing, SiTU, and checkpoint weight mapping
• supports the official full configuration and the public eight-layer
  inference-optimization/Kimi-K3-0.40B-MXFP4 development checkpoint
• supports TP=1 with MXFP4 routed experts through the offload-family backends
• makes hybrid KDA/MLA KV allocation layer-aware
• bounds CUDA graph batch sizes against the available MoE cache slots
• makes the offload copy plan size from cache capacity instead of expert count
• loads custom Hugging Face tokenizers with trust_remote_code=True and adds
  the required tiktoken dependency
• selects naive request caching for Kimi until linear-attention state snapshots
  can support radix reuse safely

Validation

• 61 focused model, weight-loader, KV-pool, MoE, graph, and tokenizer tests
  passed while CUDA was healthy
• 22/22 config, loader, graph-batch, and tokenizer tests passed again after the
  benchmark host lost GPU access
• exact checkpoint audit loaded 204 resident tensors with matching shapes and
  all seven routed-expert banks from revision
  ef3902c318fb8e13c3507e26055656e687fdfe38
• OpenAI-compatible streaming generation ran end to end on an RTX 3090 through
  32 simultaneous requests

Raw metrics, methodology, host/checkpoint metadata, and the disclosed host
failure are preserved in the
benchmark artifact.

The benchmark uses the 0.40B development checkpoint only as a functional test.
Its throughput is not presented as a proxy for the full 2.8T model. During a
sustained repetition, the live-swapped NVIDIA driver lost the passed-through
GPU; completed samples are retained, the interrupted sample is not reported,
and the failure is documented in the artifact.

Scope

• text generation only
• tensor parallel size 1
• Kimi routed experts must be MXFP4
• hybrid radix reuse is intentionally disabled for Kimi pending KDA state
  snapshot support

This contribution was directed by the operator and implemented/tested with
Codex assistance. The implementation was manually checked against the official
configuration and checkpoint keys.

Full-checkpoint qualification evidence

The full moonshotai/Kimi-K3 checkpoint passed an end-to-end OpenAI-compatible qualification on 4× NVIDIA H200. The versioned record includes the exact server arguments, 21m30s cold start to first /v1/models, warmup and streaming measurements, and a 1/2/4/8/16-request concurrency sweep with 31/31 successful requests.

• Human-readable H200 qualification record
• Machine-readable H200 measurements
• docs/models.md Supported Models entry

The branch was rebased on upstream main (bd372b6). The Kimi, generic offload, GPT-OSS, and graph batch-size regression suite passes: 59 tests.