perf: union rdeps queries across changed modules into one subprocess

[rdark]

Partial fix for #335

CalculateImpactedTargetsInteractor.queryTargetsDependingOnModules previously spawned one bazel query rdeps(//..., @<repo>//...) subprocess per changed-module-matching canonical repo. A single changed module can substring-match thousands of repos in a bzlmod workspace, and each subprocess pays ~2s of JVM + bazel-client-connect overhead serially.

Union all matched repos into a single rdeps(//..., @@a//... + @@b//... + ...) query. Bazel computes the reverse-dep graph of //... once regardless of how many patterns are in the union, so runtime collapses from N × (startup + analysis) to 1 × (startup + analysis); the N-1 eliminated subprocesses are the bulk of the saving.

• No command-line length concern: BazelQueryService.runQuery writes queries via --query_file, so the query string is arbitrary size.
• Failure semantics: a single try/catch wraps the unioned query; on failure, fall back to marking all workspace targets impacted. The previous outer catch-all is removed - audit confirmed every throwable call now has a tight try/catch around it, and the broad catch was silently swallowing errors.
• Per-module log line preserves "module X → matched N repos: ..." attribution so as to not break logging contract
• Tests: adds testUnionsRdepsAcrossChangedModules with two changed modules to assert the single-query invariant.