Fix stale embedding client during knowledge base indexing

[stellanhou]

Summary

This PR fixes a stale embedding client issue during knowledge base initialization.

When users change the active embedding provider/model, diagnostics can read the updated configuration correctly, but LlamaIndex indexing may still reuse a previously cached embedding client. As a result, indexing can fall back to the old/default embedding provider.

This change:

• resets the cached embedding client when configuring LlamaIndex embedding settings
• refreshes LlamaIndex settings before knowledge base initialization

Test

Manually verified that knowledge base initialization uses the active SiliconFlow embedding configuration instead of falling back to OpenAI text-embedding-3-large.

The issue was reproduced locally when switching from the default OpenAI embedding configuration to SiliconFlow BAAI/bge-m3.