fix: prevent session context leakage by clearing Ollama KV cache on flush

[flysheep-ai]

Summary

LibreChat currently does not clear the Ollama model's server-side KV cache when flushing the application cache, leading to context bleeding between separate chat sessions. A user reported receiving responses containing context from a previous BAML test run after starting a new chat (issue #14669).

This fix extends the cache flushing script (config/flush-cache.js) to also send a request to Ollama's /api/clear endpoint (if configured), ensuring the model's context cache is emptied alongside application caches.

Changes

• Add flushOllamaCache() function that sends a POST to Ollama's /api/clear endpoint
• Detect Ollama host from OLLAMA_HOST environment variable (defaults to http://localhost:11434)
• Add proper error handling: timeouts and connection failures log warnings but never block the flush
• Integrate the model cache clearing step after existing cache flushing operations

Validation

• Dry-run mode (--dry-run) shows the endpoint that would be called without making the request
• Verbose mode (--verbose) logs the exact endpoint being contacted
• If Ollama is unreachable, a warning is logged and the flush continues normally
• If Ollama is not in use (default state for most deployments), the /api/clear call simply fails gracefully

Risks

• If the Ollama server is down or unreachable, a warning is logged but the flush continues — model cache clearing is best-effort
• Clearing the model KV cache is a global operation affecting all active model contexts, which is expected during a cache flush

Closes #14669