Fix MPS device support for Apple Silicon Macs

Problem

StableDiffusion.setup() used a binary device check (cuda or cpu) that ignored Apple's MPS backend. On Apple Silicon, self.device resolved to cpu, while ComfyUI's model management moved the UNet to MPS. This caused a device mismatch — encode_text produced CPU-resident embeddings that the MPS-resident UNet rejected at the cross-attention layer:

RuntimeError: Tensor for argument input is on cpu but expected on mps

Changes (stable_diffusion.py)

1. Device detection now includes MPS — replaced the one-liner torch.device("cuda" if torch.cuda.is_available() else "cpu") with a three-way check: CUDA → MPS → CPU.
2. Text encoder .to() captures return value — changed self.text_encoder.to(...) to self.text_encoder = self.text_encoder.to(...) so the reference always points to the materialized model, which matters for lazy/meta tensors on MPS.
3. Guard in encode_text — added self.text_encoder = self.text_encoder.to(self.device) before inference to force weight materialization, preventing MPS failures from lazy-loaded embedding weights.

No functional change on CUDA or CPU.