fix(serving): serialize SSE generator + FlashInfer plan dtype (bf16)

[drunkcoding]

General serving-infra fixes surfaced while productizing chunked-prefill (see docs/superpowers/reports/2026-08-30-productization-verdicts.md). These are independent of the chunked feature itself.

• api_server_v2: persist a single shielded next(stream) future across the 0.1s disconnect-poll so concurrent next() no longer raises ValueError: generator already executing (this hung all streamed requests). Also log + set UNHEALTHY when the engine loop dies instead of hanging silently.
• attention_backend: pass q_data_type/kv_data_type (bf16) to the FlashInfer prefill+decode plans — the float16 default caused dtype of q bfloat16 does not match q_data_type float16, which silently killed the async engine loop on Qwen3.
• kv_cache.set_block_store: normalize cuda device index in the geometry check.

Verified: whole-prefill now generates ('The capital of France is' -> ' Paris...'). NOTE: chunked-prefill itself remains blocked by a separate KV block-budget sizing issue (logical 10940 > physical 9725) documented in the report — not addressed here.

[drunkcoding]

Recommended first merge. Of the productization PRs (#187, #188, #189, draft #190), this one is the lowest-risk and most broadly beneficial:

• Both fixes are general serving-infra bugs, independent of chunked-prefill — they fix the OpenAI server's streamed-completion path and the FlashInfer plan dtype for any bf16 model on FlashInfer, not just this feature branch.
• Concretely: (1) eliminates ValueError: generator already executing that hung every streamed request, and (2) fixes the float16-vs-bf16 plan mismatch that silently killed the async engine loop. Also surfaces engine-loop crashes as UNHEALTHY instead of hanging.
• No API/behavior change beyond making streaming + bf16 FlashInfer actually work; verified whole-prefill generates end-to-end.

Suggest landing this first (or cherry-picking the two commits to main), then the feature-specific PRs on top.