Implemented progressive </think> logit-bias control as a llama.cpp sampler patch

[tls5657]

Hi, I built a small llama.cpp patchset for more graceful reasoning-length control.

Instead of hard-cutting reasoning at token N, this patch progressively increases the logit of the </think> token as thinking gets longer.

What it adds:

• per-request think_control JSON in the OpenAI-compatible request body
• server-side validation and token resolution
• a custom sampler inserted into the native sampler chain
• 3 phases:
  • stage1: gentle bias
  • stage2: stronger bias
  • plateau: strongest bias

This was motivated by the same kind of problems discussed in:

• Feature Request: Graceful reasoning budget termination. Avoid mid-sentence cutoff. #20632
• Reasoning effort experiments #12339
• Feature Request: add per-request "reasoning" options in llama-server #13272
• llama-server needs support for reasoning_effort #20408

Repository:
https://github.com/tls5657/think-control

If there is interest, I can also share benchmark results and tuning notes for Qwen-style reasoning models and VLM + structured-output workloads.