Fix: echo reasoning_content back for DeepSeek V4 multi-turn tool calls (fixed the Deepseek V4 thinking problem)

[Psy-lzh]

Fix: echo reasoning_content back for DeepSeek V4 multi-turn tool calls

Two layers in the CLI silently dropped assistant thinking blocks before they could reach build_chat_completion_request, breaking multi-turn tool-using conversations with DeepSeek V4 (which mandates that prior reasoning_content be echoed back per its thinking-mode protocol).

1. AnthropicRuntimeClient::consume_stream did not accumulate
   ThinkingDelta into pending state, and never emitted
   AssistantEvent::Thinking on ContentBlockStop. As a result the
   session never contained a ContentBlock::Thinking.

2. convert_messages mapped ContentBlock::Thinking -> None, dropping
   thinking blocks even if they had been persisted.

build_chat_completion_request already serializes
InputContentBlock::Thinking to the reasoning_content field for
DeepSeek V4 (gated on model_requires_reasoning_content_in_history),
so no provider-side changes are needed.

Verified end-to-end with deepseek-v4-flash via a local SSE proxy:
multi-turn tool-using conversations now succeed where they
previously failed with "400 The reasoning_content in the thinking
mode must be passed back to the API."

[Psy-lzh]

Thanks a lot for the sharing by the way