VLM: wire performance instrumentation and logits output

[stikves]

Wire performance metrics and logits output for the VLM inference path, matching the LLM path behavior.

Changes

• Report prompt throughput (prefill t/s) and generation throughput in the performance summary
• Wire --print-logits for VLM: shows top-5 token probabilities per generated step
• Wire --save-logits for VLM: saves top-K logits to JSON (same format as LLM path)
• Make TokenLogits and TopLogitEntry properties public (cross-module access)

Sample output

$ llm-runner --model <vlm-bundle> --image photo.jpg --prompt "What is this?" --max-tokens 10

Generating...
The image features a small kitten sitting on

Performance Summary:
==================================================
Model Load: 15299.1ms
Prompt:     1028.3ms, 590 tokens, 573.8 tokens/sec
Generation: 480.8ms, 9 tokens, 18.7 tokens/sec
Total:      24.479s
==================================================

With --print-logits:

Generating...
  logits top5: [450]=26.812 [910]=25.406 [512]=24.422 [319]=22.609 [739]=22.109 The
  logits top5: [1967]=25.562 [9088]=19.609 [7623]=18.797 [15373]=18.734 [17739]=17.516 image
  ...

With --save-logits /tmp/vlm_logits.json:

{
  "tokens": [{
    "token_id": 450,
    "incremental_text": "The",
    "top_logits": [
      {"token_id": 450, "incremental_text": "The", "logit": 26.8125},
      {"token_id": 910, "incremental_text": "This", "logit": 25.406}
    ]
  }]
}

Test plan

• Build passes
• VLM shows correct prompt/generation t/s in performance summary
• --print-logits displays top-5 logits per token during VLM generation
• --save-logits produces valid JSON with top-K entries
• Verbose output (--verbose) shows full timing breakdown table
• Verify no regression on text-only LLM path

[carinapeng]

Thank you for taking this on @stikves !

We seem to have implemented the runner-level design ( in runVLMInference, call setPromptTokenCount(vlmTokens.count) and wrap the prefill + generation) #70

I proposed engine level here as well, seems to me it could be a more sustainable design because if we instrument CoreAISequentialVLMEngine then metrics can work for any caller, that's how we do it for text engines as well

I wonder if that'd be a better design to be more generic?