Remove unused HashSet allocations in HashJavaNames

[jonathanpeppers]

The hashes32 and hashes64 HashSet<T> instances in TypeMappingReleaseNativeAssemblyGenerator.HashJavaNames were populated every iteration but never read from -- pure wasted allocations during type map generation.

This was found while analyzing a binlog for GenerateTypeMappings task performance.

[Copilot]

Pull request overview

This PR improves type map generation performance by removing unused per-call HashSet<T> allocations in TypeMappingReleaseNativeAssemblyGenerator.HashJavaNames, where the sets were populated but never read.

Changes:

• Remove unused HashSet<uint> and HashSet<ulong> allocations from HashJavaNames.
• Keep generating/storing both 32-bit and 64-bit Java name hashes on each TypeMapJava entry.

[jonathanpeppers]

This change shouldn't affect anything, there are a couple flaky CI failures we can ignore:

Analysis
3 of 4 failures are performance tests (PerformanceTest.cs) that exceeded timing thresholds. These are commonly flaky on shared CI agents due to variable machine load. The margins were relatively small (e.g., 3268ms vs 3000ms threshold for Build_XAML_Change(False)).
1 failure is a Gradle binding test (BindLibraryWithMultipleGradleVersions) that failed during project build with AGP 8.5.0, Gradle 8.7, compileSdk 34. Unrelated to these changes.
Infrastructure issue: PowerShell crashed in MSBuild+Emulator 1 job post-test step with ManagedPSEntry+StartupException / procargs errno 5. This is an agent-level issue unrelated to the PR changes.