Enhance worker thread performance and add SocketWorker class

[jsilvanus]

This pull request introduces a new benchmarking suite for the worker pool and makes several improvements to worker communication, task queue efficiency, and embedding result handling. The key changes include adding micro-benchmarks for threads and processes, improving how embedding results are transferred and reshaped, and optimizing the pending task queue for better performance under load.

Benchmarking and Testing:

• Added a new micro-benchmark script bench/micro-bench.js to compare thread and process worker pool performance, along with mock worker scripts for both process and thread modes. [1] [2] [3]

Worker Communication and Embedding Handling:

• Enhanced the thread worker script to transfer embedding results as Float32Array buffers for zero-copy efficiency, including shape metadata for proper reconstruction.
• Updated the worker pool to automatically reshape received embedding buffers (with shape info) back into number[][] arrays before resolving tasks, ensuring a consistent public API regardless of worker implementation.

Performance and Robustness Improvements:

• Replaced the pending task array in WorkerPool with a custom FIFOQueue class to avoid performance issues with large queues and frequent Array.shift() calls. [1] [2] [3] [4]

New Worker Implementation:

• Added a new SocketWorker class in src/socket-worker.js to support connecting to a socket-based model server using an NDJSON protocol, exposing an EventEmitter interface for compatibility with the pool.