tensor update somehow breaks with cuda

[mehrhardt]

When updating to the most recent version of odl I receive the following error message from some code that worked well before:

NotImplementedError: no corresponding tensor space available for space FunctionSpace(IntervalProd([-177.3321 , -177.3321 , -128.984375], [ 177.3321 , 177.3321 , 128.984375]), out_dtype='float32') and implementation 'cuda'

I suppose @kohr-h knows where it breaks. It would be good if either 'cuda' is supported with tensors or it falls back to the old behaviour but breaking old code is no good.

[kohr-h]

@adler-j is working on a fix. The change was too fundamental, falling back to old behavior is not an option. The only way forward is to build some quick index helper that flattens multi-indices (that's the main breaker), but it won't be particularly fast. I'm also working on a replacement.

Sorry for the intermediate "outage", but it's kinda planned.

[adler-j]

It should be mentioned that the most likely fix to this is to merge #1231, which should act as a replacement. It is highly likely that odlcuda may even be considered obsolete after that (since cupy is in most senses strictly better).