MedicalNetPerceptualSimilarity: Add multi-channel

[SomeUserName1]

Fixes #7567 .

Description

MedicalNetPerceptualSimilarity: Add multi-channel support for 3Dvolumes.
The current version of the code in the dev branch already largely supports that besides the following:
medicalnet_* require inputs to have a single channel.
This PR passes the multi-channel volume channel-wise to the networks and concatenates the resulting feature vectors.
The existing code takes care of averaging over channels and spatially.

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[KumoLiu]

The implementation appears to be done correctly and it successfully introduces multi-channel input support for PerceptualSimilarity. It effectively allows each channel to be processed independently, which is a substantial addition.
However, I'd like to suggest the consideration of a specific use-case scenario. The new implementation seems to be most useful under situations where independent channel processing is beneficial. Conversely, in the final computation stage, results.sum(dim=1, keepdim=True) consolidates the multi-channel perception into a single scalar evaluation. This approach works well under single modality situations.
For cases involving multi-modal inputs, it might be worth considering the separate evaluation of each modality. As a suggestion, perhaps we could add an optional flag that switches between combined evaluation and individual modality evaluation. This could potentially enhance the generalization capabilities of the implementation and provide users the flexibility to choose based on their specific use case dynamics.

[SomeUserName1]

Will adjust accordingly on monday, such that the return types are scalar or 1D array/tensor respectively.

[SomeUserName1]

I tried to adapt it for RadNet and ResNet too, however as they expect 3 channel input and always output 2048 features, it's non trivial to treat the images channel-wise here.
At least i can not think of a trivial way to disentangle the effect that each of the 3 channels has on the 2048 features.

So I'd stick to MedicalNet for a channel-wise perceptual loss.

Will squash the commits as soon as LGTY

[SomeUserName1]

@KumoLiu only the blossom-ci check hasn't passed. Can we merge?

[KumoLiu]

Thanks for the quick update, few comments inline.

[KumoLiu]

Hi @SomeUserName1, I noticed that you add me review the PR again.

In case you didn't noticed the comment here. Looks a little confuse to me here. Could you please help explain it more?
#7568 (comment)
Thanks.

[KumoLiu]

Thanks for the quick update, overall looks good to me.
Only a small comment, will merge this ASAP.

[KumoLiu]

/build