This PR adds an optional --metrics-json=<path> flag to examples/llm to export structured timing metrics for reproducible performance tracking.

What it adds

New CLI flag: --metrics-json=<path>

• JSON metrics output with:
• model
• backend
• seqlen
• tokenizer_load_ms
• weights_load_ms
• compile_ms
• generation_ms (non-interactive --prompt mode)

Why

This enables benchmark automation and easier regression detection across commits/backends by providing machine-readable run metrics instead of relying only on logs.

Behavior notes

No behavior change when --metrics-json is not provided.
Metrics are primarily intended for non-interactive runs (--prompt).

Example output snippet

{
  "model": "hf://meta-llama/Llama-3.1-8B-Instruct",
  "backend": "cuda_fa3",
  "seqlen": 2048,
  "tokenizer_load_ms": 118,
  "weights_load_ms": 8420,
  "compile_ms": 15640,
  "generation_ms": 972
}