[codex] Add contextual precision grouping

[cat0825]

Summary

• Add an optional group_by hook to ContextualPrecisionMetric for scoring overlapping retrieval chunks as retrieval units.
• Preserve current per-context scoring when group_by is not provided.
• Keep ungrouped contexts as singleton units when group_by returns None.
• Use raw context text for verdict generation while allowing group_by to inspect the original retrieval context item, including RetrievedContextData.source.

Closes #2594

Context

PR #2659 currently remains open but conflicting. This PR applies the same current-main direction as a focused opt-in grouping implementation with deterministic tests.

Validation

• poetry run pytest tests/test_metrics/test_contextual_precision_grouping.py -q
• OPENAI_API_KEY= poetry run pytest tests/test_metrics/test_contextual_precision_metric.py tests/test_metrics/test_contextual_precision_grouping.py -q
• poetry run black --check deepeval/metrics/contextual_precision/contextual_precision.py tests/test_metrics/test_contextual_precision_grouping.py
• poetry run ruff check deepeval/metrics/contextual_precision/contextual_precision.py tests/test_metrics/test_contextual_precision_grouping.py

Note: running the existing OpenAI-backed contextual precision tests with the local OPENAI_API_KEY present fails in this environment before metric execution because the machine has a SOCKS proxy configured and socksio is not installed. The command above clears OPENAI_API_KEY so those integration-style tests follow the repository skip rule while the new deterministic tests run.

[vercel]

@cat0825 is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.

[cat0825]

Update: I opened #2737 as the source-first follow-up to #2594/#2669.

I am treating this draft as superseded rather than continuing the group_by API direction. The new PR keeps the RetrievedContextData.source approach discussed by maintainers, adds source-grouping parity for TurnContextualPrecisionMetric, and includes deterministic tests/docs around source granularity and first-seen rank behavior.