llms: preserve reasoning_content in MessageContent for tool calling round-trip

[krus210]

fix: preserve reasoning_content in tool calling round-trip for deepseek-reasoner

Closes #1478

Summary

• Fix reasoning_content being lost during conversation round-trip when using deepseek-reasoner with tool calling
• The DeepSeek API requires every assistant message to include reasoning_content -- without it, the API returns HTTP 400: Missing reasoning_content field in the assistant message
• This is the same class of bug that was found in the Python LangChain library

Problem

The reasoning_content field was correctly parsed from API responses (PR #1121) but could not be sent back in subsequent requests because:

1. MessageContent struct had no ReasoningContent field -- users couldn't attach reasoning content to assistant messages in conversation history
2. The MessageContent -> ChatMessage conversion in the OpenAI provider didn't transfer reasoning content

Changes

File	Change
llms/generatecontent.go	Add ReasoningContent string field to MessageContent struct
llms/openai/openaillm.go	Copy mc.ReasoningContent to msg.ReasoningContent when converting AI messages
llms/marshaling.go	Update MarshalJSON/UnmarshalJSON to include reasoning_content (with omitempty)
llms/marshaling_test.go	Add test cases for reasoning_content marshal/unmarshal/roundtrip
llms/openai/internal/openaiclient/chat_test.go	Add test for ChatMessage with reasoning + tool_calls

Usage

After this fix, users can preserve reasoning_content in the tool calling round-trip:

// Step 1: Get response with reasoning + tool calls
resp, _ := llm.GenerateContent(ctx, messages, llms.WithTools(tools))
choice := resp.Choices[0]

// Step 2: Build assistant message preserving reasoning_content
assistantMsg := llms.MessageContent{
    Role:             llms.ChatMessageTypeAI,
    Parts:            []llms.ContentPart{choice.ToolCalls[0]},
    ReasoningContent: choice.ReasoningContent,  // <-- now possible!
}

// Step 3: Send back with tool result -- no more HTTP 400
resp2, _ := llm.GenerateContent(ctx, []llms.MessageContent{
    userMsg,
    assistantMsg,
    toolResultMsg,
}, llms.WithTools(tools))

Test plan

• Added unit test: ChatMessage marshal/unmarshal with ReasoningContent + ToolCalls
• Added unit test: MessageContent marshal/unmarshal with reasoning_content
• Added unit test: round-trip MessageContent with reasoning_content + tool calls (JSON and YAML)
• All existing tests pass (go test ./llms/...)
• Verified with live DeepSeek API: two-step tool calling flow completes successfully

Notes

• The fix is minimal and backward-compatible -- ReasoningContent defaults to empty string and is omitted from JSON when empty
• The OpenAIFunctionsAgent also doesn't propagate reasoning_content through its scratchpad (agents/openai_functions_agent.go). This is a separate issue that can be addressed in a follow-up PR
• DeepSeek API docs on thinking with tools: https://api-docs.deepseek.com/guides/thinking_with_tools