feat(deepseek-v2-lite): add mixed-request serving gate with per-request KV ownership

[CAICAIIs]

Summary

This PR adds the first greedy mixed-request serving gate for DeepSeek-V2-Lite EP2.

• add MixedRequestScheduler for explicit admit / reject / defer handling
• keep one owned DecodeCache per active request
• retire requests independently on EOS, length, rejection, error, or sink failure
• batch decode only when active requests share the same decode position; otherwise fall back to single-row decode
• extend the EP2 E2E to cover same-position batching, single-row fallback, and invalid-request isolation
• update the DeepSeek-V2-Lite status docs to keep the claim boundary explicit

Validation

• cargo fmt --all --check
• git diff --check
• cargo test --offline --release -p openinfer-deepseek-v2-lite --features deepseek-v2-lite --lib -- --nocapture
• host-staged EP2 E2E
• NCCL EP2 E2E with an updated NCCL runtime
• HF / host-staged / NCCL exact compare

Result: all_token_text_exact

Scope

Not claimed here: vLLM parity, production EP readiness, sparse dispatch, or HTTP continuous batching.

Closes #281

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ca281080ea

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Avoid cloning full KV caches on each batched decode

When two or more active requests share a decode position, this snapshots every active DecodeCache before every generated token just to support fallback. These caches are the full f32 per-layer KV history and grow with context; for DeepSeek-V2-Lite that is hundreds of MB per token for modest 8-row batches, so the intended same-position mixed-serving path can be dominated by host copies or hit memory pressure even when batched decode succeeds. Avoid taking full-cache clones on the hot path, or make fallback use a smaller rollback strategy.

Useful? React with 👍 / 👎.