HTTP client connection leak in embedders causes file descriptor exhaustion

[choutos]

Summary

The AsyncOpenAI, Ollama, and VoyageAI embedders create HTTP clients that are never explicitly closed. Over time, connections accumulate in CLOSE_WAIT state and eventually exhaust available file descriptors.

Symptoms

• Too many open files errors after prolonged operation
• Connections stuck in CLOSE_WAIT state (visible via ss -tan or netstat)
• Gradual file descriptor exhaustion under sustained embedding workloads

Root Cause

The embedder classes instantiate HTTP clients (e.g., httpx.AsyncClient via the OpenAI SDK, or direct httpx usage in Ollama/VoyageAI) but lack cleanup logic to close them when the embedding loop exits.

Proposed Fix

See PR #917, which:

• Adds a cleanup() method to the Embedder base class
• Implements cleanup() in affected embedders to close underlying HTTP clients
• Calls cleanup() in Executor.run()'s finally block
• Reuses client instances instead of creating new ones per request (Ollama, VoyageAI)