feat: add native continuous batching and chunked prefill for LLM runtime

[ajmalrasi]

Summary

Adds native continuous batching for the LLM runtime, allowing up to two active
text-generation sequences to share an existing batch-two TensorRT engine.

The implementation is built around persistent sequence-local runtime state,
rather than concurrent calls to the existing single-request API.

What changed

• Added stable two-slot sequence ownership with generation-tagged handles.
• Added persistent request-local KV, recurrent, convolution, prompt and output state.
• Added selected-row native execution so one or both physical slots can run safely.
• Added bounded chunked prefill with a qualified 128-token policy.
  • Resumed final chunks use 64–128 tokens.
  • Remainders of 129–191 tokens are partitioned safely.
• Added a decode-first native scheduler with:
  • bounded FIFO admission,
  • round-robin prefill progress,
  • safe-boundary slot reuse,
  • cancellation and failure propagation.
• Added independent request controls for sampling, output limits, EOS, stop strings,
  thinking, logit bias and logprobs.
• Added native ticket/channel integration for independent HTTP and SSE responses.
• Added streaming/non-streaming handling, disconnect cleanup, tool/reasoning parsing,
  usage accounting and overload responses.
• Added finite startup CUDA graph captures for slot views {0}, {1} and {0,1}.

Validation

• Host regression suite: 146 passing tests.
• Native scheduler/state tests passed with ASan/UBSan and ThreadSanitizer coverage.
• Jetson Orin validation confirmed exact serial/concurrent output equality for the
  qualified execution paths.
• Chunked-prefill validation covered supported boundaries through 6,144-token prompts.
• HTTP gates covered staggered requests, mixed streaming, cancellation, EOS/stops,
  logprobs, tool calling, reasoning, overload and cleanup.
• Performance qualification on a batch-two Orin engine:
  • aggregate output throughput: 22.54 → 33.97 tok/s (1.51×)
  • singleton latency: within 4.3% of the legacy path
  • staggered second-request TTFT: 2.048 → 0.202 seconds

Explicit non-goals

• No cross-request context/KV reuse or prefix caching.
• No speculative decoding or MTP.
• No arbitrary chunk boundaries; the raw short-tail 64+64+1 path remains unsupported.
• Scope is the validated single-rank, text-only, batch-two runtime.