[Feature] Add Unsloth Studio UI value to tune the llama-server --fit-target flag for squeezed extra performance.

[xyehya]

Current implementation using --fit on uses dynamic approximation to identify how many layers llama-server can fit into GPU while leaving a default buffer of around 1GB unused --> forcing more layers into CPU
In some situations where: 1- VRAM is tight (e.g 16GB VRAM) and 2- the GPU is used purely for compute and not graphics (the nvidia_drm module is not loaded) setting --fit-target to 128/256 will 1.enable all layers on the GPU or in the best case 2. fit all layers + the sane KVCache configured into GPU --> at least 10x performance difference in pp and pg..

[aiSynergy37]

Opened PR #4882 implementing this request.

What it adds:

• Chat Settings UI control for GGUF fit_target (Auto, 64, 128, 256, 512)
• API field plumbing (fit_target) across load request/response + status
• backend command wiring: passes --fit-target when --fit on is active
• runtime state restore from /api/inference/status
• regression tests in studio/backend/tests/test_native_context_length.py

[xyehya]

@aiSynergy37 need help testing on my side? I am still getting an issue with forcing a specific context length behind a certain limit. Even though I get a warning (Exceeds estimated VRAM capacity (8,960 tokens). The model may use system RAM.) applying it still falls back to 8,960 tokens of KV cache. Although at this ctx my vram usage is 14GB/16GB (RTX 5080). The backend doesn't seem to respect the UI values.