Torch backend returns tensor of different shape and data type then input

[biagio-lunit]

First of all, thank you for implementing this library, it is useful to have different stain normalization algorithms with a simple pip install.

Describe the bug
Macenko normalizer changes the image shape and dtype after normalization,
which requires additional steps in the dataloader which is unexpected and unnecessary.

To Reproduce

H, W, C = 100, 100, 3
target = torch.rand(C, H, W).to(torch.float32)
query = torch.rand(C, H, W).to(torch.float32)
normalizer = torchstain.normalizers.MacenkoNormalizer(backend="torch")
normalizer.fit(target)
query_norm = normalizer.normalize(query)[0]
assert query_norm.shape == query.shape, f'{query_norm.shape} == {query.shape}'
assert query_norm.dtype == query.dtype, f'{query_norm.dtype} == {query.dtype}'

the output is

AssertionError: torch.Size([100, 100, 3]) == torch.Size([3, 100, 100])
AssertionError: torch.int32 == torch.float32

Expected behavior
Image before and after normalization should be the same shape and type.