Get label probabilities

[KrisThielemans]

• add `get_label_probabilities() (Get label probabilities #8 <- extract labelfields #2)
  • add Act.all_labels
  • expose Act
• add trim_zeros_ROI()
• cache resolution and PET activity class (add capability of using different resolution/shape in get_mmr_* #6 <- allow specifying different resolution #4)
• make toPetMmr() return ndarray
• update example images
• update dependencies

This PR (#8) fixes #2

Also this PR supersedes:

• closes fix handling of arguments in get_mmr_fromfile #5 (fixes get_mMR_from_file ignore parameters? #3) already merged
• closes add capability of using different resolution/shape in get_mmr_* #6 (fixes allow specifying different resolution #4)

[KrisThielemans]

This PR has a problem. skimage.transform.resize is used for interpolation as follows

brainweb/brainweb/utils.py

Lines 445 to 447 in 202eb40

	arr = resize(arr, new_shape,
	order=1, mode='constant', anti_aliasing=False,
	preserve_range=True).astype(dtype)

However this does not give "probability maps", and there's no guarantee they'd add to 1 at all. In fact, I get values slightly negative or larger than 1 at edges of the masks (which I don't understand).

This seems therefore only correct when using outres=brainweb.

[casperdcl]

Sorry somehow missed the notification about this! Will review and merge in <24h