Fix MaxRecursionDepth error, Memory leaks and move to CUDA12

[chaithyagr]

We had some issues of MaxRecurssionDepth, which was fixed and on extensively testing, I saw a memory leak in gpuNUFFT (fixed in chaithyagr/gpuNUFFT#23). Some more memory leaks from cuFFT comes from from CUDA11.8: https://forums.developer.nvidia.com/t/nvc-cufft-memory-leak/279476/3

To fix this, we moved to using CUDA12.1, which is default version which comes with torch.

Also have a bunch of other miscellaneous tasks:

1. Added support for binder button in docs.
2. Fixed a minor bug in smaps estimation which didnt allow int values for blurr_factor.

[paquiteau]

Great Work ! moving to cuda12 is really nice. We should probably make a release after this.

I have some minor comments/suggestions to address before that though ;)

[paquiteau]

so whats the expected behavior ? one needs to put docs_build to skip the building of the documentation ? seems weird to me...

[chaithyagr]

Its weird, when I added this, it is building .. I am lost but it works :P

[paquiteau]

Let's clear this then, We should build the doc:

• In PR to master, when the docs_build keyword is present.
• On every commit to master (aka after merge)

https://docs.github.com/en/actions/writing-workflows/choosing-what-your-workflow-does/contexts#github-context

I think we should use github.head_ref (and the contains)

[chaithyagr]

The ref aspect is fine. The containts part doesnt work. I am trying to understand whats broken. For some reason it does not trigger even when I use docs_build

[chaithyagr]

Okay this is broken. For now, we dont check docs_build and also style. (Cause github.event.head_commit is empty if the event is triggered with pull_request I will open an issue (#176) to fix this, but its not easy :P
We can proceed with current PR for now