[EvilFreelancer]

What

load_model(device=None) picks cuda if available, otherwise cpu, so on a machine with an Intel GPU (PyTorch xpu backend) GigaAM silently runs on the CPU unless the caller passes device="xpu". The model itself already works on XPU as it is: GigaAMASR.forward autocasts on the model's own device type, and nothing else is CUDA-specific.

This PR:

• adds _default_device(): cuda, then xpu (when torch.xpu exists and reports a device), then cpu;
• extends the existing flash_attn fallback from "disable on CPU" to "disable on anything but CUDA", since flash_attn is a CUDA-only library;
• adds tests/test_device.py (9 tests, no model download): the default-device order including a torch build without torch.xpu, an explicit device being kept, and the flash_attn policy per device type.

Tested

On an Intel Arc A770 (xe driver, torch 2.10.0+xpu, Ubuntu 24.04 container):

• pytest tests/test_device.py: 9 passed;
• gigaam.load_model("v3_e2e_rnnt") without device loads onto xpu:0;
• long-form transcription of a 712 s Russian recording (40 VAD chunks, batched RNNT decoding, fp16 encoder) gives the same text word for word as an RTX 5060 Ti with torch 2.10.0+cu128, in about the same time (14.5 s against 13.7 s).

Nothing changes for CUDA or CPU users.