Next release - Tensorflow to Pytorch

[pgunn]

The next release of Caiman will switch the tensorflow parts of the codebase out for pytorch implementations. This will make packaging for Windows easier again (particularly the conda packages - Windows users on conda have been stuck on older releases for awhile), and ease the way forward in packaging in other ways. We'll be able to make some more substantial changes after this lands.

This work is thanks to Manuel Paez, who has seen success with using onnx tooling to help translate the models.

As part of this release, I intend to add a reliable marker in the hdf5 serialisations of OnAcid and CNMF objects indicating which datatype it came from, something simply and reliably usable to distinguish them, as well as an API call that uses the marker (although a significant use of the marker will be for other software that doesn't necessarily import the caiman codebase, that may not even be written in Python, to distinguish).

Because of the packaging implications, the next release will likely be 1.13.0 rather than continuing the 1.12 line