[codex] Speed up differential coverage output generation

[0xalpharush]

Summary

This is a performance follow-up to #29. It keeps the new relcov non-inferiority statistics unchanged, but removes avoidable work in write_differential_coverage_outputs().

• Skip oversized non-combined campaigns before sorting/writing normalized showmap files.
• Reuse one DifferentialCoverage object per analyzed campaign across relscore, relcov, and verdict generation.
• Compute relcov tables and per-seed relcov samples from cached approach unions/trials instead of repeatedly constructing one-trial ApproachData objects.
• Add regression tests for skipped campaign materialization and redundant DifferentialCoverage construction.

Validation

• uv run --with-requirements analysis/requirements.txt python -m unittest analysis.tests.test_differential_coverage analysis.tests.test_benchmark_report
• git diff --check origin/main..HEAD