feat: support tool-specific call limits

[tian44200]

Summary

Feature: Add tool-specific call limits
(Currently, we only have tool_call_limit that applies to the total count across all tools, not for individual tools.)

Why this feature: The query "search for Sherlock Holmes in the knowledge base" might trigger the search_knowledge tool up to 8 times in a single run. But we might only want it once in some cases. See more details in link to discord discussion.

Changes:
This PR only covers the following for now:

• Custom tools → added call_limit for Function type
• search knowledge (default tool) → added search_knowledge_call_limit for Agent(..)

This PR does not cover:

• Toolkit and functions inside Toolkit (can be extended easily with the current implementation, but not sure about DX: add call limit for each tool inside or the whole toolkit? or later?)
• Dict
• Other default tools (read_chat_history, update_knowledge ...)

To be discussed:
As mentioned in Discord, the DX need to be discussed. The current implementation is just a suggestion and can be changed:

# with decorator
@tool(call_limit=3)
    def limited_func() -> str:
        return "test"

# as argument
func = Function(name="test_func", call_limit=5)

# for search knowledge (default tool)
agent = Agent(
    model=OpenAIChat(id="gpt-4o-mini"),
    knowledge=loaded_knowledge_base,
    search_knowledge=True,
    search_knowledge_call_limit=1,
    ...
)

Small Note: I noticed that continue_run/acontinue_run actually resets tool_call_limit. Not sure if this is intended. This PR also followed the same implementation logic, but if not, the issue can be handled in a later PR or this if needed.

Type of change

• Bug fix
• New feature
• Breaking change
• Improvement
• Model update
• Other:

---

Checklist

• Code complies with style guidelines
• Ran format/validation scripts (./scripts/format.sh and ./scripts/validate.sh)
• Self-review completed
• Documentation updated (comments, docstrings)
• Examples and guides: Relevant cookbook examples have been included or updated (if applicable)
• Tested in clean environment
• Tests added/updated (if applicable)

---

Additional Notes

The actual implementation follows the logic of the existing tool_call_limit logic so it should be safe.
I also tested on my personal project.
For search_knowledge_call_limit=1 and query="search for reinforcement learning and tabular reinforcement learning":

 [
{'id': 'd3e1857c1f6c4cf0bc7e9a4cb3c7fb1b', 'type': 'function', 'function': {'name': 'search_knowledge_base', 'arguments': '{"query": "reinforcement learning"}'}}, 
{'id': 'ad3a3b6703ac4022ba4f1ad5cd766c75', 'type': 'function', 'function': {'name': 'search_knowledge_base', 'arguments': '{"query": "tabular reinforcement learning}'}}]}, 
{'role': 'tool', 'content': "Call limit reached for search_knowledge_base. Don't try to execute this tool again.", 'tool_call_id': 'ad3a3b6703ac4022ba4f1ad5cd766c75'}, 
{'role': 'tool', 'content':'[\n  {\n    "content" ...', 'tool_call_id': 'd3e1857c1f6c4cf0bc7e9a4cb3c7fb1b' }, ...
]