fix: Mistral Medium 3.5 reasoning – use reasoning_effort=high and parse ThinkChunk streaming format

Mistral Medium 3.5 with thinking enabled requires reasoning_effort="high" (not the
legacy thinking:{type:"enabled",budget_tokens:...} block). When streaming with
reasoning_effort="high", delta.content becomes a list of ThinkChunk objects
(type="thinking", containing a  array) followed by TextChunk objects
(type="text"). Only after the thinking phase does delta.content revert to a plain
string. The old code only checked delta.reasoning_content / delta.reasoning and
missed all thinking tokens entirely.

Changes:
- _callMistral: replace thinking:{...} with reasoning_effort:"high"
- _callMistral streaming loop: handle Array.isArray(d.content) ThinkChunk/TextChunk
  list format; fall back to plain-string path for non-reasoning models

Author: Android-PowerUser