🐛 Describe the bug

torch.cdouble (complex128) tensors can be moved to the mps device without error via .to(device), but subsequent operations fail. Direct conversion to complex128 on the mps device is correctly blocked by a TypeError.

Reproduction

import torch
device = torch.device("mps")

# 1. Incorrectly allows move to MPS
x = torch.randn(2, 2, dtype=torch.cdouble, device="cpu").to(device) 
print(x.dtype) # torch.complex128 on mps

# 2. Crash on operation
# RuntimeError: Undefined type ComplexDouble
y = x * x 

# 3. Correctly blocks conversion (Inconsistent with above)
# TypeError: Trying to convert ComplexDouble to the MPS backend...
z = torch.randn(2, 2, device=device).to(torch.cdouble)

Expected Behavior

.to(device) should raise a TypeError when attempting to move a complex128 tensor to the MPS backend, as it is unsupported. Error messages should be consistent across conversion methods.

Versions

PyTorch: 2.10.0
OS: macOS (Apple Silicon, MPS enabled)
Device: MPS

cc @malfet @ezyang @anjali411 @dylanbespalko @mruberry @nikitaved @amjames @kulinseth @DenisVieriu97 @jhavukainen @aditvenk