Fix Bifrost source-less model symbol parsing

[DavidBakerEffendi]

Description

Update the UsageBench Bifrost adapter to parse the producer's tagged authored and virtual model-location variants. Valid source-less model symbols no longer abort an entire search_symbols response, while authored symbols still require exact source identity before they can select a benchmark declaration.

Key Changes:

• Deserialize authored { path, symbol, range } and virtual { uri, range } model locations without weakening malformed authored-location rejection.
• Exclude virtual model locations from source-backed declaration matching and retain exact path, line, kind, and name checks for authored candidates.
• Add regressions for the original mixed authored/source-less response, a malformed authored location, and a same-name authored symbol in another file.

Validation:

• cargo test: 273 passed.
• cargo run -- validate benchmarks/cases: 59 benchmark case files validated.
• Public Bifrost 724e596e308191d1b5c9349bfb0585e2ab135497, with JAVA_HOME='': all 18 formerly failing Java/Scala cases passed with no diagnostics. Java completed 23/23; Scala completed with 14 passes and one existing expected failure, with no hard failures or runner errors.
• cargo fmt --check still reports only pre-existing formatting drift in src/promotion.rs; the affected src/runners/bifrost.rs is formatted.
• cargo clippy --all-targets -- -D warnings still reports nine pre-existing warnings outside this change's logic; the adapter-owned warning was fixed. No warnings were suppressed.

Touch Points:

• src/runners/bifrost.rs