feat: wire MPS (Apple Silicon) device into fine-tuning wizard

[SahilKumar75]

Context

Issue #26 tracks MPS support generally. The fine-tuning wizard (added in #33) defaults to device_map="auto" in both trainer/qlora.py and the inference cache in app/api.py. On Apple Silicon, bitsandbytes 4-bit quantization (used by QLoRA) is not supported — attempting to run QLoRA on MPS will either fall back to CPU or error out.

What needs to happen

1. Detect device at wizard start — call GET /api/gpu on page mount in FinetuneState and store the backend (cuda / mps / cpu).

2. Disable QLoRA on MPS — if backend == "mps", grey out the QLoRA technique button with tooltip: "QLoRA requires CUDA. Switch to LoRA for Apple Silicon." Auto-select LoRA instead.

3. Set correct dtype in loader — trainer/loader.py should pass torch_dtype=torch.float16 and skip BitsAndBytesConfig when MPS is detected.

4. Inference endpoint — app/api.py::infer hardcodes torch_dtype=torch.float16 and device_map="auto". On MPS this needs .to("mps") instead of device_map="auto" (which is CUDA-only).

Acceptance criteria

• Wizard detects MPS and auto-selects LoRA on Apple Silicon machines
• QLoRA button is visually disabled with an explanatory tooltip on MPS
• Training runs without error on M1/M2/M3 Mac using LoRA (no quantization)
• Test chat inference works on MPS after training