[None][feat] publish native V2 KV cache events

[alec-flowers]

Summary

• add a top-level strict KVEventsConfig with vLLM-compatible enablement, publisher, endpoint, replay, buffering, HWM, queue-size, and topic fields
• reuse TensorRT-LLM V2's existing mutation hooks and iteration flush, replacing the attention-DP gather callback with local per-rank event conversion and publishing
• publish vLLM-compatible msgpack batches over the three-frame ZMQ protocol with sequence numbers and base_port + rank endpoint behavior
• emit complete blocks from the maximum attention-window lifecycle, suppress trailing partial blocks and their removal events, and ignore native created/storage-tier update events for this proof of concept
• support attention-DP and conventional TP without KV-event object collectives; reject unsupported PP/CP topologies
• add bounded background queues, null-publisher construction-only mode, counters, and idempotent shutdown
• update the LLM-args telemetry golden manifest for the new public nested configuration

The paired Dynamo consumer is available in ai-dynamo/dynamo#12167.

Validation

• two focused native-event tests passed in the TensorRT-LLM 1.3.0rc21 runtime, covering the no-gather callback, vLLM wire decoding, partial-block filtering, sequence numbering, and shutdown/port reuse
• LLM-args golden-manifest generation check passed
• Python compilation and git diff --check passed
• local Qwen3-0.6B TensorRT-LLM smoke test delivered a stored-event batch containing two full blocks to the real Dynamo KV router; shutdown counters reported one published batch, zero dropped batches, and zero KV-event allgathers

The rc21 benchmark implementation remains on a separate branch; this draft is the clean forward-port against current main.