google-common: Support Gemini 2.0 Thinking model

[afirstenberg]

Privileged issue

• I am a LangChain maintainer, or was asked directly by a LangChain maintainer to create an issue here.

Issue Content

Note this is NOT about general Gemini 2.0 compatibility. This is specifically about the "gemini-2.0-flash-thinking-exp" model (and related "reasoning" models).

There are some issues

• The "reasoning tokens" are available in the output, but they are a different text part than the final result
  • So part[0] contains the reasoning and part[1] contains the final result
  • If using API version "v1alpha" (at least on AI Studio), there is a "thinking" attribute which is available to indicate which part is the reasoning and which part is final
  • The current logic tries to merge text parts together. This is, arguably, undesirable. At the very least - it breaks some of the tests.
  • We need to see how this works with streaming.
  • We need to discuss the best way to handle this in general - not just for Gemini.
• Currently, this model doesn't support tools.

My thinking about handling the reasoning tokens:

• For compatibility with other models - they should be removed from normal output.
• They should be included in the response_metadata as a "reasoning" attribute, or something similar.
• We should have a filter that can, optionally, put them back together.
• Should any of this be controlled with a parameter?
• I don't think this needs to be a new class.