fix: adjust VRAM for 24K context based on actual usage

Based on error output, model uses ~17.5GB (not 15GB estimated).
- Llama: 85% VRAM for 24576 context (3GB KV cache)
- BGE: 6% VRAM (reduced to fit)
- Total: 91%

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

services/vllm/config_bge.yaml

@@ -2,6 +2,6 @@ model: BAAI/bge-large-en-v1.5
 host: "0.0.0.0"
 port: 8002
 uvicorn-log-level: "info"
-gpu-memory-utilization: 0.08
+gpu-memory-utilization: 0.06
 dtype: float16
 task: embed

services/vllm/config_llama.yaml

@@ -2,7 +2,7 @@ model: meta-llama/Llama-3.1-8B-Instruct
 host: "0.0.0.0"
 port: 8001
 uvicorn-log-level: "info"
-gpu-memory-utilization: 0.80
+gpu-memory-utilization: 0.85
 max-model-len: 24576
 dtype: auto
 enforce-eager: false