fix(utils): exclude parameters with default values from required

[Ghraven]

Summary

convert_function_to_tool currently lists every signature parameter in required, even ones that have a default value. Under both the JSON Schema and OpenAI tool schema spec, a parameter with a default is optional and should not be in required.

Repro

from ollama._utils import convert_function_to_tool

def get_weather(city: str, units: str = 'celsius') -> str:
    """Get the weather.
    Args:
        city: The city.
        units: Temperature units.
    """
    return ''

tool = convert_function_to_tool(get_weather).model_dump()
print(tool['function']['parameters']['required'])

Before: ['city', 'units']
After: ['city']

In practice this confuses some tool-calling models into always passing every argument (often with the literal default repeated back), and it doesn't match what the upstream OpenAI / Anthropic-compat endpoints expect.

Fix

1. Collect parameter names whose inspect.Parameter.default is not inspect.Parameter.empty.
2. After pydantic generates the schema, remove those names from schema['required'].
3. Hardened the existing Union[X, None] block to check membership before calling .remove() so the two paths don't double-remove.

The pydantic model is still built from __annotations__ + __signature__, so introspection metadata is unchanged.

Tests

• Added test_function_with_default_values_not_required — covers a mix of typed defaults including bool and str.
• Added test_function_with_all_defaults_has_empty_required — covers the all-optional case.
• All 12 tests in tests/test_utils.py pass locally.

$ pytest tests/test_utils.py -q
............                                                             [100%]
12 passed in 0.24s

Happy to adjust naming, split commits, or fold the regression test into the existing parametrized case if that would fit the codebase better.