ENH dwidenoise: Support >4D data

[Lestropie]

Initially mentioned in #2743, but giving its own Issue for purposes of listing.

Currently, when I run dwidenoise on multi-echo fMRI data, I have to concatenate echoes along the volume dimension, denoise, and then separate into separate echo files based on slicing along the volume axis. It would be nicer if instead the input image could be 5D: 3 spatial dimensions, timepoints, echo. That's a simple mrcat call, and then using multi-file numbered image could be used to separate back across echoes.

[Lestropie]

Better would be to actually implement the tensor PCA approach, since it should be applicable for any acquisition that can be represented as an ND (N>4) dataset.