Release v0.22.0

Added

• EvaluationOptions.Thinking — the reasoning mode of every evaluator's verdict call (Faithfulness, Relevancy,
  Answerability, AnswerCoverage), passed to CompletionOptions.Thinking.
• OpenAICompatibleCompletionService.ReasoningSwitch — how CompletionOptions.Thinking reaches an OpenAI-compatible server.
  The service ignored Thinking, so the Off that chunk enrichment and (now) the evaluators ask for never reached a reasoning
  model on such a server. OpenAICompatibleReasoningSwitch.ChatTemplateKwargs sends chat_template_kwargs
  (enable_thinking / thinking) for On / Off — vLLM, SGLang, llama.cpp (GPUStack included); None (the default) sends
  nothing, which keeps OpenAI itself working. Also AddFluxImproverWithOpenAI(…, reasoningSwitch, …).

Changed

• Breaking — evaluators ask for no reasoning by default (ThinkingMode.Off). They sent no thinking mode, so a grader
  that reasons on its default could spend MaxTokens on hidden reasoning and return no verdict — a scoring failure, not a
  score. Same judgement as EnrichmentOptions.Thinking. Migration: to keep the model's own default, set
  new EvaluationOptions { Thinking = ThinkingMode.Auto } (and raise MaxTokens if it reasons).

Dependencies

• Re-pinned sibling package(s) LMSupply.Generator 0.128.0 -> 0.129.0, LMSupply.Generator.Onnx 0.128.0 -> 0.129.0.
• Re-pinned sibling package(s) LMSupply.Generator 0.129.0 -> 0.131.0, LMSupply.Generator.Onnx 0.129.0 -> 0.131.0.