Add max_retries to AgentSkillsToolset

[kaxil]

Stacked on #74379. Until that merges, this diff also shows its commit; the changes here are the last two commits.

pydantic-ai-skills gives its tools a budget of one correction, and that toolset-level value overrides the agent's retries. So a model that named two missing skill resources in a row failed the run with UnexpectedModelBehavior: Tool 'read_skill_resource' exceeded max retries count of 1, even with agent_params={"retries": {"tools": 3}}. AgentSkillsToolset had no way to change that.

AgentSkillsToolset now takes max_retries with the same meaning it has on the SQL, hook, DataFusion and object storage toolsets:

• None, the default, uses the agent's tool retry budget.
• An integer gives the skills tools a budget of their own.

It works by setting each tool's budget in AgentSkillsToolset.get_tools, resolved the same way as AirflowToolset._get_tool_max_retries. Nothing is forwarded into pydantic-ai-skills, so the fix does not depend on how that library handles the value. With the agent's default retries of one, nothing changes for existing Dags.

A real-run test covers three cases:

• With the default budget, the second refused read still fails the run.
• With the agent's retries raised to 3, the run finishes. This case fails without the change.
• An explicit max_retries=2 wins over the agent's value.

The restricted Agent Skills example now sets max_retries=3. The guide and the retry-budget section of the toolsets index describe the parameter, with both outcomes re-captured from a real run.