fix(memory): Implement engine & model cache lifetime management with configurable idle TTL and auto-eviction #97

Description & Evidence

During live system operation, terminating LocalLLMServerManager (PID 37952) resulted in an immediate reclaim of 12.29 GB of host system RAM and ~870 MB of Windows Non-Paged Kernel Pool memory, even though the application's reported private working set was only ~1.85 GB.

Collected System Metrics

Root Cause

1. Unbounded Model Residency in Host & Driver Memory: When backend engines (Ollama, ComfyUI, Forge, Kokoro) load models or memory-map large GGUF/safetensors files, host memory pages and CUDA/driver host-pinned buffers remain allocated indefinitely.
2. Missing Cache Lifetime & Idle TTL Policies: VramOrchestrator currently only attempts to unload models reactively when another workload explicitly demands VRAM (EnsureVramForImageGenerationAsync(), EnsureVramForComfyUiAsync()). It does not enforce any idle expiration policy, TTL countdown, or keep-alive window.
3. Ollama Keep-Alive Unmanaged: Requests sent to Ollama default to retaining the model in memory indefinitely (keep_alive = -1 or engine default) unless explicitly told otherwise.

Proposed Changes

1. Configurable Model Cache Lifetimes: Add a user setting for model idle lifetime (ModelCacheIdleTimeoutMinutes, e.g., 5m, 15m, 30m, 1h, Never).
2. Idle Eviction Watchdog: Implement a background service in VramOrchestrator that monitors the last-used timestamp of loaded models and triggers an unload (/api/generate with keep_alive: 0 for Ollama, /free for ComfyUI) once the idle TTL expires.
3. Graceful Engine Hibernation: Provide an auto-stop or hibernation trigger for dormant background child engines after extended inactivity.