Feature request: OpenAI-compatible and Gemini API embedding support

[ysys143]

Problem

Local GGUF embedding works well for privacy/offline use, but has friction:

• Requires downloading a ~300MB+ model on first run
• Slow on CPU; needs Metal/CUDA for reasonable throughput
• Limited to models that fit in local memory

There's no way to use cloud embedding APIs (OpenAI, Gemini, Ollama remote, etc.) without forking.

Proposed solution

Add three environment variables to activate API-based embedding as an alternative to local GGUF:

QMD_EMBED_API_URL=   # base URL — presence activates API mode
QMD_EMBED_API_KEY=   # API key
QMD_EMBED_API_MODEL= # model name

API type auto-detected from URL — no extra config:

• googleapis.com → Gemini (batchEmbedContents)
• anything else → OpenAI-compatible (OpenAI, Ollama, LM Studio, etc.)

Local mode continues to work exactly as before when QMD_EMBED_API_URL is unset.

Benchmark (63 Korean+English chunks, M3 Max)

Model	Per chunk	Dims	Cost
embeddinggemma-300M (local/Metal)	72ms	768	$0
text-embedding-3-small	16ms	1536	$0.020/1M tokens
text-embedding-3-large	13ms	3072	$0.130/1M tokens
gemini-embedding-001	38ms	3072	free tier

Implementation

PR #427 has a working implementation with benchmark script.