refactor: modularization and performance

[paulocoghi]

This PR is focused on better organizing the codebase and applying some initial performance and efficiency optimizations.

Architectural Changes

• Splitting of large classes: Mongo::Client and Mongo::Collection were split into their respective subfolders and multiple smaller modules like client/retryable_writes.cr, collection/find_and_modify.cr.
• Modular Unified Runner: The large spec/unified_runner.cr was split into multiple modules at spec/unified/, introducing a clean Dispatcher pattern that makes it a lot easier to expand and improve the driver.

Performance & Memory Optimizations (reduced allocations)

• Socket Streaming: BSON parsing off the TCP socket no longer allocates intermediate arrays. It uses read_fully with IO::Memory
• User-Space Pipes: GridFS now uses jgaskins/pipe instead of kernel-level IO.pipe. This has the potential to bring up to ~10x faster pipe throughput.
• Lock-Free IDs: Replaced the OS-level Mutex for tracking request_ids with a lock-free Atomic(Int32).
• GC Relief: Eliminated chained .map/.select array materialization, switching to single-pass compact_map and String.build.

Crystal 1.20/1.21 Modernization & Safety

• Zero .not_nil!: All instances of .not_nil! have been removed. The driver relies entirely on Crystal's type narrowing and raises descriptive Mongo::Errors if anomalies occur.
• Execution Contexts: Migrated to Sync::Mutex and removed all assumptions of single-thread fiber pinning, making the driver robust for Crystal 1.21 Parallel contexts.

Test Status: 324/324 Unified Test Format examples correctly passing (18m 55s execution time)