feat: add QMD_EMBED_MODEL env var for multilingual embedding support

[daocoding]

Problem

The default embeddinggemma-300M embedding model is English-centric and produces poor vector representations for CJK (Chinese, Japanese, Korean) text. Users working with multilingual document collections have no way to switch to a better embedding model without modifying source code.

Solution

Add QMD_EMBED_MODEL environment variable to override the default embedding model at runtime.

Changes

• src/llm.ts: DEFAULT_EMBED_MODEL now reads from QMD_EMBED_MODEL env var (falls back to embeddinggemma-300M for backward compatibility)
• src/llm.ts: getDefaultLlamaCpp() passes QMD_EMBED_MODEL to LlamaCpp config when set
• src/llm.ts: Add isQwen3EmbeddingModel() helper to detect Qwen3-Embedding model family
• src/llm.ts: formatQueryForEmbedding() and formatDocForEmbedding() auto-detect model family and apply the correct prompt format:
  • embeddinggemma: nomic-style task: search result | query: ... (unchanged)
  • Qwen3-Embedding: task-instruction format Instruct: ...\nQuery: ...
• src/store.ts: pass model URI to format functions for consistency between indexing and query time
• README.md: document QMD_EMBED_MODEL with usage example

Usage

# Qwen3-Embedding-0.6B is multilingual (119 languages), MTEB top-ranked at its size
export QMD_EMBED_MODEL="hf:Qwen/Qwen3-Embedding-0.6B-GGUF/qwen3-embedding-0.6b-q8_0.gguf"

# Re-embed after switching models
qmd embed -f

Notes

• Fully backward-compatible — no change when QMD_EMBED_MODEL is unset
• The reranker (Qwen3-Reranker-0.6B) already ships with QMD; this PR aligns the embedding side for users who want a consistent Qwen3 pipeline
• Only format detection logic added; no behavior change for default model