[BFCL] Use underscore_to_dot config flag for tool name compilation

[lukeinglis]

Summary

• Fixes the underscore_to_dot flag to control dot-to-underscore replacement during tool compilation, not just during evaluation
• Replaces the hardcoded ModelStyle list in convert_to_tool() with the existing ModelConfig.underscore_to_dot flag
• Passes the flag from build_handler() through to all handler _compile_tools methods

Problem

The underscore_to_dot flag in ModelConfig was only used by the evaluation checker (ast_checker.py), but the actual dot-to-underscore replacement in convert_to_tool() used a hardcoded list of ModelStyle values. This caused a mismatch: models with underscore_to_dot=False still had their function names mangled during tool compilation, leading to evaluation failures on test cases with dotted function names (e.g., math.factorial became math_factorial).

This particularly affects models evaluated via the /v1/chat/completions FC path (e.g., through proxies or gateways), where the model sees underscored tool names and returns underscored calls, but the evaluator compares against the original dotted ground truth.

We observed ~50 point drops on affected categories (e.g., multiple: 94.5% to 37%) until we worked around this in a custom handler.

Changes

• convert_to_tool(): replaced hardcoded ModelStyle list with underscore_to_dot parameter (default False for backward compatibility)
• build_handler(): passes config.underscore_to_dot to handler constructor
• All _compile_tools implementations: forward the flag to convert_to_tool()

Backward Compatibility

Models with underscore_to_dot=True in their ModelConfig (e.g., OpenAI, Mistral, Anthropic) continue to have dots replaced, same as before. The behavior change only affects models with underscore_to_dot=False (or new/custom models), which will now correctly preserve dotted function names.

Fixes #1091