feat(code_gen): add per-language subset metrics for livecodebench-x

[rodboev]

compute_metrics() currently groups by difficulty only. livecodebench-x benchmark rows include a target_language field (de, es, fr, ja) as a top-level data row field (adjacent to verifier_metadata, not inside it), but the field was never declared on CompCodingVerifyRequest — Pydantic silently dropped it — and therefore never reached CompCodingVerifyResponse or compute_subset_metrics. As a result, per-language pass@k values never appear in rollouts_aggregate_metrics.json.

This PR fixes that by mirroring the existing difficulty pattern:

• Adds target_language: Optional[str] = None to CompCodingVerifyRequest so Pydantic preserves the top-level field through routing
• Adds target_language: Optional[str] = None to CompCodingVerifyResponse
• Reads body.target_language in verify() and propagates it through all three return sites via **body.model_dump()
• Calls compute_subset_metrics(tasks, "target_language", ...) alongside the existing difficulty call
• Confirms get_key_metrics() already handles the new prefixes generically — no change needed there
• Adds focused tests for the new metric keys and field propagation

Closes #1171

Checklist

• Targeted tests pass: uv run pytest resources_servers/code_gen/tests/test_app.py -x -v
• No changes to benchmark data files, config YAML, or nemo_gym/reward_profile.py
• DCO sign-off on all commits

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.