Streaming SSE responses from OpenAI-compatible providers (e.g. DeepSeek
thinking-mode) can include reasoning_content deltas alongside content and
tool_calls. Previously the parseStreamResponse Delta struct ignored
reasoning_content, so assistant reasoning was silently lost.
Fix: add ReasoningContent to the stream Delta struct, accumulate it in
parallel with content, and include it in the returned LLMResponse.
This intentionally does NOT change the existing DeepSeek replay rules in
filterDeepSeekReasoningTurn. Those rules correctly enforce the documented
vendor contract: reasoning_content must be preserved for tool-interaction
turns but is optional (ignored) for plain assistant turns.
Test: TestProviderChatStream_ParsesReasoningContent covers SSE streaming
with interleaved reasoning_content, content, tool call deltas, and usage.
e0f5ccff2d