Content field is empty in streaming response after tool call in MCP client

Bug description
When using the Spring AI MCP client for handling tool calls, we've observed inconsistent behavior between streaming and non-streaming modes regarding the content field in the request payload.

1. Streaming Mode (stream: true):

   {
     "messages": [
       {
         "content": "What data does the database contain?",
         "role": "user"
       },
       {
         "content": "",  // 👈 Content is unexpectedly empty
         "role": "assistant",
         "tool_calls": [
           {
             "id": "...",
             "type": "function",
             "function": {"name": "list_tables"}
           }
         ]
       },
       // ... subsequent tool response
     ],
     "stream": true
   }
   

2. Non-streaming Mode (stream: false):

   {
     "messages": [
       {
         "content": "What data does the database contain?",
         "role": "user"
       },
       {
         "content": "\n<think>\n ... \n</think>\n\n",  // Content contains intermediate thought steps
         "role": "assistant",
         "tool_calls": [
           // ... same tool call structure
         ]
       },
       // ... subsequent tool response
     ],
     "stream": false
   }
   

Environment

• Spring AI version: 1.0.0-M6
• Java version: 17
• MCP API version: 0.7.0-SNAPSHOT

Steps to reproduce

1. Send a tool call request with stream: true
2. Observe empty content field in assistant message
3. Compare with non-streaming request where content contains text content

Expected behavior

• Both streaming and non-streaming modes should maintain consistent payload structure
• The content field should either:
  1. Contain intermediate thought steps in both modes, or
  2. Be omitted in streaming mode while preserving tool call metadata

Start by reproducing the streaming and non-streaming MCP client requests described in the issue, then trace the Spring AI MCP client code that builds assistant messages after a tool call. Compare how content and tool-call metadata are serialized in both paths; done means the chosen content behavior is consistent and covered by a regression test.