fix(finetuning-asr): auto-detect device and attention implementation in LoRA inference

[voidborne-d]

Problem

finetuning-asr/inference_lora.py hardcodes attn_implementation=\"flash_attention_2\" and splits dtype as bfloat16 vs float32 on cpu-only. Two concrete symptoms for non-CUDA users running a fine-tuned LoRA:

• On Apple Silicon (--device mps): from_pretrained(..., attn_implementation=\"flash_attention_2\") blows up at model load — flash attention 2 requires CUDA.
• On Intel XPU or CPU: same failure.
• The --device argparse option has no choices=, so typos silently make it through to .to(device) with confusing tracebacks.

The sibling batch-inference demo at demo/vibevoice_asr_inference_from_file.py already solves this with an auto path. This PR brings the LoRA inference script to parity.

Changes

• load_lora_model() now accepts attn_implementation and threads it through from_pretrained. MPS load-then-move comment added inline.
• main() gains a --attn_implementation flag (flash_attention_2|sdpa|eager|auto, default auto). When auto, CUDA + importable flash_attn → flash_attention_2, otherwise falls back to sdpa with a log line.
• --device now restricts to {cuda, cpu, mps, xpu, auto} and defaults to the best available (cuda > xpu > mps > cpu), same pattern as vibevoice_asr_inference_from_file.py.
• dtype is torch.float32 for cpu/mps/xpu (required), torch.bfloat16 for cuda.

Behaviour for CUDA users with flash_attn installed is unchanged.

Verification

Local smoke test with load_lora_model stubbed out (no GPU needed):

=== device=cpu ===
Auto-detected attention implementation: sdpa
load_lora_model called with device=cpu dtype=torch.float32 attn=sdpa

=== device=mps ===
Auto-detected attention implementation: sdpa
load_lora_model called with device=mps dtype=torch.float32 attn=sdpa

=== device=cuda (no flash_attn) ===
flash_attn not installed, falling back to sdpa
Auto-detected attention implementation: sdpa

=== device=cpu --attn_implementation eager ===
(no auto-detect print; explicit eager flows through)

python finetuning-asr/inference_lora.py --help reflects the new flag/choices.