Optimize PMTiles memory usage and integrate shared caching

[groldan]

Refactor the PMTiles implementation to reduce memory footprint and leverage the new shared caching infrastructure.

• Zero-Allocation Optimizations:
  • Optimized HilbertCurve to use primitive math and binary search for zoom level lookup, eliminating object allocation during tile ID conversion.
  • Replaced ByteBufferInputStream with CompressionUtil utilities that stream directly from SeekableByteChannel through BoundedInputStream and Channels, avoiding intermediate buffer copies.
  • Refactored PMTilesEntry to an interface with value-based implementations (Leaf, Tile) to minimize object overhead for directory entries.
  • Updated PMTilesDirectoryImpl to use binary search on the packed ByteBuffer for entry lookups instead of deserializing all entries into objects.
• Shared Caching Integration:
  • Integrated PMTilesReader and PMTilesVectorTileStore with CacheManager to use shared cache resources.
  • Implemented DirectoryCache and VectorTileCache as views of the shared cache, ensuring efficient eviction and global memory limits.
• API Improvements:
  • Updated PMTilesReader API to use TileIndex consistently and expose getTileId for direct access.
  • Removed redundant internal buffering logic in favor of the shared ByteBufferPool and direct channel operations.

This changeset significantly reduces heap pressure when reading large PMTiles archives and improves concurrent read performance.

Refactor caching to use shared Caffeine cache with SubCache support

Introduces a centralized caching infrastructure to optimize memory usage across multiple RangeReader instances.

• Shared Cache Architecture: Added Cache, CacheManager, and SubCache abstractions in io.tileverse.cache. Multiple logical caches now share a single underlying Caffeine cache instance, allowing for global eviction policies and shared weight limits across the application.
• RangeReader Integration: Refactored CachingRangeReader to use RangeReaderCache, which acts as a view (sub-cache) of the shared global "tileverse-rangereader-cache", keyed by the source identifier.
• HTTP Optimizations: Updated HttpRangeReader to use a shared CachedThreadPool for HttpClient.
• API Updates: CachingRangeReader.Builder now accepts a CacheManager.
• Improve ByteBufferPool cache locality, stats, and direct buffers clean up

This architecture prevents memory fragmentation and ensures that total cache usage respects global limits rather than per-instance limits.