refactor: introduce canonical tool naming convention and update evalsets & model generators to reflect this change.

[omkargaikwad23]

Summary

The changes centralize tool name normalization into a new utility module, allowing the trajectory matcher to remain generator-agnostic while ensuring that tool identities are preserved across different agent platforms

Key Changes

• Canonical Naming Convention: Standardized MCP tool names to the server__tool format (using double underscores) to distinguish between tools with the same name across different servers.
• Normalization Utility: Added evalbench/generators/models/tool_naming.py to handle specific formatting differences from Claude Code (mcp__), Gemini CLI (mcp_), and Codex CLI.
• Agnostic Scoring: Refactored TrajectoryMatcher to remove per-generator normalization logic, enabling simpler string-based comparisons of pre-normalized trajectories.
• Dataset Migration: Updated expected_trajectory entries in evaluation sets (e.g., Cloud SQL and Bigtable datasets) to reflect the new canonical format.
• Improved Testing: Added comprehensive unit tests for the naming helpers and updated trajectory matcher tests to verify the new standard.

[IsmailMehdi]

/gcbrun