fix: GGUF local model GPU crashes and rough edges

[Dak0verflow]

Summary

• Enable flash attention (and quantized KV cache) for GGUF GPU inference instead of forcing it off.
• Add a repetition penalty to the GGUF sampler chain.
• Add a local model picker with live import progress and friendly display names.
• Raise the GGUF import cap from 3 blobs to 8.
• Simplify the system prompt for local models under ~2B parameters.

Why

• GPU crash: Standard (non-flash) attention's Vulkan matmul dispatch hits a ggml-vulkan descriptor-set accounting bug on some GPUs (reproduced on a Mali-G715 / Tensor G4, no coopmat/int-dot support) — a hard GGML_ASSERT abort that kills the process outright. Because it's a signal, not a C++ exception, the existing Vulkan-error-to-CPU fallback (built to handle things like VK_ERROR_DEVICE_LOST) never sees it and can't recover. Flash attention routes through a different kernel path that avoids the bug entirely. llama.cpp requires quantized KV cache whenever flash attention is enabled, so both changed together.
• Repetition penalty: the sampler chain had no repetition/frequency penalty at all (top_k -> top_p -> temp -> dist). Small local models (sub-2B) would reliably degenerate into repetitive or self-referential loops without one.
• Model picker + import progress: the "Model file" field showed the raw internal blob path (blob_<uuid>.gguf) instead of the imported file's name, and importing a multi-GB model gave zero visible feedback during the copy — it looked hung. Switching models required re-importing every time since there was no way to pick from what was already imported.
• Import cap: maxBlobCount was hardcoded to 3 regardless of the actual 24GB aggregate storage budget, an arbitrary extra ceiling that got in the way of trying more than 3 models.
• Tiny-model system prompt: the local-mode system prompt includes a detailed JSON tool-call protocol (control markers, schemas, an approval-capability workflow) on top of the base prompt. Models under ~2B parameters (tested: Qwen3-0.6B, Qwen3-1.7B) reliably lose the thread on this and degenerate into one-word non-answers (e.g. a bare "Done") instead of answering or calling a tool correctly. They're not capable enough to use the protocol reliably in the first place, so skipping straight to a plain conversational prompt below that size loses nothing.

Test Plan

• Reproduced the GPU crash on-device (Pixel 9 Pro XL / Tensor G4) with both a standard-quant model and a ternary-quant model; confirmed the exact GGML_ASSERT failure via llvm-addr2line/llvm-objdump against the built .so, then confirmed the fix resolves it for both on GPU.
• Verified the repetition penalty change against a sub-1B model that reliably looped without it.
• Manually exercised the model picker: imported multiple models, confirmed friendly names and live progress, confirmed switching between already-imported models without re-importing.
• Verified the tiny-model prompt path against two sub-2B models; confirmed two larger models (4B and 8B) are unaffected.