fix(bench): keep reports flowing without install check

[zackees]

Summary

The benchmark was failing after completing measurements because its final fbuild install --check --json metadata probe is rejected for registry-pinned non-ESP32 platforms. That made a post-measurement metadata step block publication of otherwise valid timing data.

This reads package identities from the successful timed build output instead. The report now distinguishes matched, different, and unverified stacks. Missing or inconsistent identities across any cold trial are recorded as unverified and suppress only the comparison ratio; the cold/warm timing bars still publish. ESP32-S3 compares platform, framework, and toolchain; Uno compares framework and toolchain.

Validation

• soldr cargo test -p fbuild-bench-fastled-examples --bin bench-build-comparison — 30 passed
• soldr cargo clippy -p fbuild-bench-fastled-examples --bin bench-build-comparison -- -D warnings — passed
• soldr cargo fmt --all -- --check — passed
• Parser format checked against a successful benchmark log; offline metadata failure reproduced in failed run #36345385136

The updated benchmark workflow watches this source path, so merging this PR should automatically run and publish a fresh report.