perf(server): reduce HTTP and PD streaming overhead

[WANDY666]

Serialize common chat/completion SSE deltas directly, coalesce events at PD batch boundaries, and cache per-stream tool metadata. Keep event ordering, tool arguments, logprobs/echo and usage unchanged; flush buffered output before stream errors.

Also avoid per-message receive tasks and allocating a fake token list solely for length validation. These remove CPU work on the streaming path without waiting for the next model token to flush a completed batch.

This covers only the remaining optimizations; compact forwarding stays in #1486 and #1487 (the latter needs rebasing for main's updated ObjType values). Tests: streaming/API and PD regressions (79 passed).