bug: ToolManager promises type coercion but passes strings to tools

[BhoomiAgrawal12]

Describe the bug
ToolManager claims to perform automatic type coercion, but _process_tool_call() passes LLM JSON string values directly to tools. Tools annotated with float/int receive strings and fail (often returning an error string in tool responses).

Expected behavior
If a tool parameter is annotated as float or int, values like "35.0" or "5" from LLM JSON should be coerced to 35.0 and 5 before the tool is invoked.

To Reproduce

1. Create a tool with an annotated numeric parameter (float or int).
2. Invoke it through ToolManager.call_tools() with JSON string values.
3. Observe the tool fails or returns an error response.

Minimal reproducible example:

from mesa_llm.tools.tool_decorator import tool
from mesa_llm.tools.tool_manager import ToolManager
import json
from unittest.mock import MagicMock

tm = ToolManager()

@tool(tool_manager=tm)
def set_value(agent, amount: float) -> str:
    return f"Set to {amount:.2f}"

tool_call = MagicMock()
tool_call.function.name = "set_value"
tool_call.function.arguments = json.dumps({"amount": "35.0"})  # string from LLM
tool_call.id = "tc_001"

class FakeAgent:
    model = MagicMock()

results = tm.call_tools(FakeAgent(), MagicMock(tool_calls=[tool_call]))
print(results)
# Expected: [{'response': 'Set to 35.00', ...}]
# Actual:   [{'response': "Error: Unknown format code 'f' for object of type 'str'", ...}]