Fixes

• Track streaming reasoning and final content in separate state so DeepSeek thinking boundaries no longer depend on rendered </think> text.
• Add a Thinking Block setting with collapsed, expanded, and thinking-only expanded display modes.
• Collapse thinking blocks automatically once final answer content starts when the thinking-only mode is selected.

Fixes

• Keep DeepSeek V4 Pro thinking content out of default answer output.
• Collapse rendered thinking blocks by default so reasoning is not shown before the final answer.
• Strip <think>...</think> blocks from answer copy, read-aloud, and saved Markdown output.
• Use a linear scanner for stripping thinking blocks to avoid regex backtracking on malformed model output.

Fixes
- Fix DeepSeek V4 Pro streaming thinking output being split into many separate thinking blocks.
- Consecutive OpenAI-compatible reasoning_content chunks now stay in the same <think> block while final answer content remains outside the block.