Support pydantic schemas in tools

[aymeric-roucher]

Tools can use pydantic models as their input schemas!

Builds on #1643, fixes #699

Features:

• Automatic Pydantic detection: Tools can now accept Pydantic BaseModel classes as parameter types
• Dict-to-model conversion: Automatically converts dictionary arguments to Pydantic model instances during tool execution (this is specially needed for ToolCallingAgent, where the model can only output a dictionary)
• Schema integration: Pydantic JSON schemas are processed and integrated with smolagents validation system
• Updated validation: Supports complex Pydantic features including nested models, enum constraints, optional fields, and custom validators
• Backward compatibility: All existing tools continue to work without changes

Discussion points:

• Requires adding pydantic dependency to smolagents
• This PR enables forcing the agent to return structured outputs, as requested many times! Indeed, one can now just create a custom FinalAnswer tool, for which the forward method takes any custom pydantic model as input, and returns its input. We could make this method more accessible by passing a structured_output arg to agent initialization, this would be for an upcoming PR

[aymeric-roucher]

@albertvillanova a big part of the logic in there is to handle our "nullable" parameter with Pydantic models that follow a different json schema spec.

Maybe it would be wise in a follow-up PR to handle json schemas directly rather than converting their logic to our nullable fomat
Handling json shcemas directly could use either of the solutions below:

Solution 1:
Just add a completely new validation logic, with a tool.input_schema parameter that follows the pydantic schema : each tool init could either (exclusive OR) use tool.inputs or tool.input_schema, depending on what was use for initialization.
Then after some time, tool.inputs could be deprecated to only follow the new version.

Solution 2:
Let tool.inputs be either our legacy way of describing inputs, or a json schema. This is detected dynamically, then either logic is used for validation.

[chahn]

@duguyue100 to my understanding, this PR might be a comprehensive solution to your issue described in #1708 and grll/mcpadapt#68

[duguyue100]

@duguyue100 to my understanding, this PR might be a comprehensive solution to your issue described in #1708 and grll/mcpadapt#68

@chahn Does this have anything to do with how MCPAdapt creates smolagents tools? I couldn't see it.
Or you are saying that MCPAdapt will somehow convert the input schema from MCP server to pydantic models and then create smolagents tools?

[aymeric-roucher]

@albertvillanova do you have some time for a review?