fix(bench): parse actual AVR package identities for Arduino Uno

[zackees]

Context

The Arduino Uno benchmark run 36381617967 reports fbuild at 425.886 ms cold and 47.307 ms warm, faster than the Arduino CLI (824.665/462.435 ms) and PlatformIO (884.674/630.370 ms). However, it warns: fbuild build output omitted resolved package identities on Arduino Uno. The result is correctly marked unverified and its fbuild/PlatformIO ratio is excluded from publication.

The build log's AVR resolved: output identifies avr-gcc@1.70300.191015 and arduino-avr-core@5.2.0. parse_fbuild_build_packages in bench/fastled-examples/src/build_comparison.rs currently searches for toolchain-atmelavr and framework-, so it does not recognize the names fbuild actually emits and returns an empty identity map. The existing test fbuild_avr_build_log_reports_selected_package_versions uses synthetic PlatformIO-style names, so it misses the real output format.

Proposal

Teach the benchmark parser to recognize fbuild's actual AVR resolved-package output and canonicalize its AVR compiler/core package names and versions into the same comparison fields used for PlatformIO. Keep the stack-verification gate: only publish a like-for-like ratio when both tools' canonical package identities and versions match.

Acceptance criteria

• RED → GREEN: add a focused parser test containing the exact real AVR resolved: log text from run 36381617967; demonstrate it fails with the current parser and passes after the fix, extracting avr-gcc@1.70300.191015 and arduino-avr-core@5.2.0 into canonical comparison fields.
• A same-stack Arduino Uno fixture with matching canonical package identities and versions is marked verified and publishes its fbuild/PlatformIO ratio.
• Missing identities or any package-version mismatch continue to mark the stack unverified and exclude the ratio.
• The existing AVR parsing and benchmark-report tests pass.

Decisions

• Treat this as a benchmark metadata/parser correctness bug (P2): the measured build is available, but package-identity parsing currently prevents a valid comparison from being published.
• Normalize fbuild's names into the existing comparison schema rather than special-casing Uno in report generation, so AVR identity verification remains reusable.

Related issues

• fix(bench): align ESP32-S3 toolchains and explain cold-build gap #1489 covers ESP32-S3 benchmark stack comparability and is related precedent; it is a separate platform/parser issue.