fix: evict idle models before loading reranker on low-VRAM GPUs

[Jua004]

Problem

On GPUs with limited VRAM (e.g. 2 GB GTX 960M), qmd query crashes at the reranking step:

Error: Failed to create any rerank context

The three models (generation ~1.2 GB, embedding ~314 MB, reranker ~610 MB) plus a rerank context (~960 MB) total ~3.1 GB — well beyond 2 GB. Since query expansion and embedding are already done by the time reranking starts, those models sit idle in VRAM while the reranker fails to allocate.

Fix

In ensureRerankModel(), before loading the reranker, check free VRAM against the reranker model file size + context overhead. If insufficient, dispose the generation model and embedding model/contexts — they've already completed their work and can be reloaded from disk later if needed.

• Uses statSync on the model file (already imported) so the threshold adapts if models change
• 15% headroom factor for runtime allocations
• On machines with enough VRAM, the check passes and nothing is evicted — zero impact on high-VRAM setups

Testing

Tested on GTX 960M (2 GB VRAM), driver 560.35.05, CUDA 12.6:

• Before: qmd query crashes at reranking every time
• After: full query pipeline (expansion → embedding → reranking) completes, VRAM peaks at ~1.9 GB

Relates to #275