Fix Ollama Empty TextPart Streaming

Description

This PR fixes an issue where streaming responses would stop prematurely when encountering empty TextPart responses from the Ollama model (served through OpenAI compatible API). The stream was incorrectly ending early when Ollama returned empty text parts before generating tool calls or additional content.

Changes

• Modified the stream_to_final function in agent.py to be more robust:
  • Only consider non-empty TextParts as potential final results using has_content() check
  • Allow empty text parts to be processed normally without ending the stream
  • Properly process TextPartDelta events with content
• Updated test snapshots to match the new behavior:
  • Streaming now produces more granular chunks in tests
  • Token counting is more accurate as we process the complete response

Testing

• Manually tested with Ollama backend to verify:
  • Streaming now continues properly through empty text parts
  • Tool calls work correctly in streaming mode
  • The streaming behavior is consistent with other model providers
• Updated test snapshots to reflect the new behavior:
  • Added additional text chunk ('The ') in streaming tests
  • Updated token counts (response_tokens: 5→8, total_tokens: 108→111)

Related Issue

Fixes #1292

Additional Notes

This issue primarily affected Ollama model responses, but the fix enhances streaming robustness across all providers by ensuring only meaningful content triggers a final result.

There is a slight increase in token counting observed in tests after this fix. This is expected and proper behavior - previously, empty text chunks would cause premature stream termination, resulting in incomplete token counting. The fix ensures the complete stream is processed correctly, resulting in more accurate token counts and proper chunking of the streamed response.