refactor(mcp_readability): group the scorer modules into a package

[akangsha7]

Addresses the review comment on #623: the readability scorers had grown to four modules sitting flat in scorers/, with carry-forward and CUJ coverage still to come. This lays them out the way dataset_quality is, as a pure move so #623 stays a review of its own logic.

Changes

1. scorers/mcp_readability/ (new package) — mcp_readability_scoring.py → scoring.py, mcp_fingerprint.py → fingerprint.py, mcp_style_readability.py → style_readability.py, mcp_tool_metrics.py → tool_metrics.py. Contents are unchanged apart from their own imports.
2. scorers/mcp_readability/registry.py (new) — SCORER_REGISTRY moves out of the orchestrator, so the package is the one thing a caller imports. It is a registry rather than a single top-level scorer class like DatasetQualityScorer: mcp_tool_metrics and mcp_style_readability are separately registered names that appear under scorers: in the run config and as the comparator on their summary rows, so collapsing them behind one scorer would change the config surface and the dashboard rows.
3. evaluator/mcp_readability/orchestrator.py — imports the registry instead of the scorer classes; docstring module paths updated.
4. Tests — import paths only. The patch targets are strings (scorers.mcp_style_readability.get_generator), so those were swept for separately; a stale one would silently patch nothing rather than fail.

Scorer names in run_config.yaml are unchanged, so the google3 mirror and Guitar are unaffected.

PYTHONPATH=$PWD/evalbench:$PWD uv run --with pytest pytest evalbench/test — 783 passed. bigtable_test fails with a GCP PermissionDenied, which reproduces on main and is unrelated.