feat(internals): Add ReasoningBudgetSampler support

- Add Python-backed ReasoningBudgetSampler for first reasoning-block control
- Install the sampler before probability filters to preserve forced end tokens
- Support reasoning_budget -1/0/N semantics in sampling params
- Force reasoning_budget_message + reasoning_end when the budget is exhausted
- Add manual force_reasoning_budget() at the sampling-context level
- Match llama.cpp force behavior by allowing only COUNTING -> FORCING
- Keep DONE as permanent passthrough and ignore later reasoning tags
- Support prefilled reasoning starts with reasoning_start_in_prompt
- Preserve UTF-8 boundary safety before forcing the end sequence
- Keep Python-backed custom sampler callbacks alive across C sampler usage
- Avoid shallow-copying custom_samplers when cloning sampler chains

Signed-off-by: JamePeng <jame_peng@sina.com>

Author: JamePeng