feat(llm): qmd query speedup - add environment variable overrides for faster model selection

[dgilperez]

Problem

When using qmd query for latency-sensitive applications (like AI agent memory search with 5-second timeouts), the default reranker becomes a bottleneck. For instance, on my Mac Mini M4 16GB, qmd query's full process takes ~15s

Solution

This PR allows to select faster models by adding environment variables for runtime model configuration without code changes:

• QMD_EMBED_MODEL - override embedding model
• QMD_GENERATE_MODEL - override query expansion model
• QMD_RERANK_MODEL - override reranker model
• QMD_MODEL_CACHE_DIR - override model cache directory

Priority: config object > environment variable > default

I managed to get great results with Jina Reranker v1-tiny - a 33M parameter distilled model optimized for speed, without changing defaults or rebuilding.

Example:
export QMD_RERANK_MODEL='hf:gpustack/jina-reranker-v1-tiny-en-GGUF/jina-reranker-v1-tiny-en-FP16.gguf'
qmd query "your search"

Benchmarks on Mac Mini M4 16GB:

Scenario	qwen3 (default)	jina-tiny	Speedup
Cold start (8 docs)	~15,000ms	~80ms	185x
Warm cache (8 docs)	400-440ms	40-65ms	6-10x
Full pipeline cold	~20s	~7s	3x
Full pipeline warm	~15s	~5.7s	2.6x

Quality: ~78% top-3 ranking agreement (same relevant docs, order differs slightly).

Includes:

• Unit tests for env var configuration
• README documentation with model override examples

No breaking changes. Default remains qwen3-reranker-0.6b.

[busbyjon]

👍 for this PR. I'm struggling to run this model on a CPU contrained system - and this would solve it!