docs: update DeepSeek README status section

[mvanhorn]

Summary

Updates the README's DeepSeek coverage so both supported lines are accurately described: the Supported Models table rows are refreshed, and the status paragraph now covers DeepSeek-V4-Flash (--features deepseek-v4, 8-GPU MP8 checkpoint, TileLang build requirement, greedy-only serving with explicit stop_reason for unsupported parameters, no CUDA Graph yet) and DeepSeek-V2-Lite (--features deepseek-v2-lite, 2-GPU EP path, current correctness-gate status). Each claim links to the measured-evidence docs under docs/models/deepseek-v4/ and docs/models/deepseek-v2-lite/.

Why this matters

#329 (split from the README rework tracker #122) asks that the DeepSeek section let users understand the supported model lines, feature flags, hardware expectations, serving status, and where the performance evidence lives. The current README covers only the V4 line in its status paragraph, has no V2-Lite status text, and links none of the evidence docs. Every claim in this update is sourced from the in-repo status/gate docs, and the V2-Lite text deliberately stays within the status ledger's claim boundaries (host-staged vs NCCL) rather than overclaiming production continuous batching.

Testing

Docs-only change. All referenced doc paths (docs/models/deepseek-v4/support.md, serving-baseline.md, decode-performance.md, docs/models/deepseek-v2-lite/status.md, hf-accuracy-gate.md) exist in the tree; feature-flag names match the per-crate Cargo.toml definitions.

Fixes #329

[CAICAIIs]

README.md:161 still describes DeepSeek-V4 as a single-request path and says bs>1 serving remains follow-up. That is stale for the current tree: the DSV4 scheduler now has the HTTP active-set decode path wired, and docs/models/deepseek-v4/online-throughput.md records active set 2 / decode batch 2 evidence with caveats. Please update this bullet to describe the limited active-set decode state and link online-throughput.md or http-serving-benchmark.md, while keeping continuous batching, multi-request prefill, service-level KV, and CUDA Graph as follow-up work.