Reduce llama.cpp scoring latency

[NotXf1le]

What changed

• Batch llama.cpp tokenization with a bounded fallback, and overlap vision capability checks with tokenization.
• Remove a redundant image-prefix retokenization request.
• Avoid extra logprob probes by default; retain exact scoring of missing choices with probeMissingLogprobs: true.

Why

These changes reduce HTTP round trips during choice scoring. The new default can assign zero probability to choices outside the returned top logprobs; callers that need their exact scores can opt in to additional probes.

Checks

• npm run check — 166 tests passed
• npm pack --dry-run
• Live RTX llama.cpp checks of default and opt-in scoring