Rename FitParams to MemProfile

The struct, runner, cache, and on-disk cache directory are now named after
what they hold (a per-model memory profile) rather than the tool that
produces them today (llama-fit-params). Also drops the 'fit' prefix from
CatalogEntry.fitResidentBytes, leaving residentBytes.

Fix Swift concurrency warnings in ModelManager and FitParams

Update llama.cpp to b8797

Fix ctxBytesPer1kTokens for Gemma 4 models (account for SWA pattern)

Suppress llama-server's automatic model discovery from HF cache

b8609's router mode auto-scans the HF cache and lists every GGUF it finds, not just models in models.ini. This causes the web UI to show models that don't appear in the app's own UI.

Switch Qwen 3.5 quantized builds from UD-Q4_K_XL to Q4_K_M for consis…

…tency with other model families

Improve models folder settings row layout

Show the path inline next to the label with truncation for long paths. Replace the text "Reset" button with a compact restore (↺) button that only appears when using a non-default directory. Both buttons use controlSize(.small) for a tighter fit.

Improve context length info text to explain memory relationship

Update llama.cpp to b7772

Update memory budget formula to totalRAM × 0.75 − 2GB and align with …

…--fit-target