fix(protocol): forward sampling masks for chat completions

[nanjiangwill]

Summary

• add return_sampling_mask to ChatCompletionRequest so typed chat routing preserves the SGLang extension
• cover enabled and default values in the existing SGLang extension serde tests
• verify the OpenAI router forwards the field to a mock worker
• initialize the recently added return_indexer_topk field in an existing GenerateRequest test fixture so the routing test target compiles

Why

Miles session rollouts call /v1/chat/completions. The generate protocol already preserves return_sampling_mask, but the chat protocol did not, so Serde discarded the field before forwarding the request to SGLang.

Tests

• cargo test --test spec_test test_sglang_extension_fields
• cargo test --test routing_tests test_openai_router_forwards_return_sampling_mask

Both focused test targets passed on the patched tree.