fix(pipeline): Preserve dtype in from_pipe() instead of defaulting to float32

[Mr-Neutr0n]

Summary

Fixes from_pipe() to preserve the source pipeline's dtype instead of defaulting to float32.

Problem

When using from_pipe() to create a new pipeline from an existing one, the dtype was not preserved:

pipe = StableDiffusionPipeline.from_pretrained(..., torch_dtype=torch.float16).to("cuda")
print(f"Before: {pipe.dtype}")  # torch.float16

i2i = StableDiffusionImg2ImgPipeline.from_pipe(pipe)
print(f"After:  {i2i.dtype}")   # torch.float32 (WRONG!)

This caused:

• Memory usage to double (from 2.6GB to 5.2GB in the example)
• Slower inference due to float32 computation

Solution

Changed the default value for torch_dtype from torch.float32 to pipeline.dtype:

# Before
torch_dtype = kwargs.pop("torch_dtype", torch.float32)

# After  
torch_dtype = kwargs.pop("torch_dtype", pipeline.dtype)

Users can still override with torch_dtype=torch.float32 if needed.

Fixes #12754