Perf: Use SPARQL VALUES for model ID restrictions

[mdorf]

Summary

• Replace large disjunctive ID filters with SPARQL VALUES clauses on AllegroGraph, Virtuoso, and GraphDB.
• Retain the existing FILTER implementation for 4store, which accepts VALUES syntax but does not apply the restriction.
• Preserve percent-encoded spaces in RDF URIs and make empty ID collections a no-op.
• Add focused coverage for backend behavior, empty IDs, URI encoding, and 500-ID batches.

Implements #191.

Verification

• Focused query-builder tests: 5 runs, 20 assertions
• test:where: 26 runs, 472 assertions
• Full test suite: 181 runs, 3,159 assertions
• RuboCop: no offenses

Performance

Measured from a local API process (Ruby 3.2.10) against the staging AllegroGraph instance containing production-equivalent data. The comparison used Goo development at cc806c8 and this branch at cb0b87b.

For each exact 500-ID SPARQL query, the benchmark used 5 warmups followed by 50 alternating measured requests per revision over persistent HTTP connections. The before/after result bindings were identical.

Workload	Query bytes (before → after)	Median (before → after)	p95 (before → after)
UBERON classes	29,361 → 26,873 (-8.5%)	190.8 ms → 186.3 ms (-2.4%)	206.1 ms → 200.9 ms (-2.5%)
NCIT classes	36,520 → 34,032 (-6.8%)	238.2 ms → 233.4 ms (-2.1%)	272.5 ms → 286.7 ms (+5.2%)

An end-to-end UBERON API benchmark (3 warmups, 20 alternating measured requests per revision) was effectively flat: median response time was 732.1 ms before and 736.9 ms after.

The measured benefit is therefore a consistent 6.8–8.5% reduction in generated query size and about a 2% improvement in query median latency for these workloads. Tail latency was mixed, and this benchmark does not demonstrate a material end-to-end API speedup.