feat: remove extra arch + refactor unit tests workflow

[bugra-yilmaz]

CUDA 12 is not compatible with some of the older GPU architectures, namely:

Allowed values for this option: 'all','all-major','compute_50','compute_52',
'compute_53','compute_60','compute_61','compute_62','compute_70','compute_72',
'compute_75','compute_80','compute_86','compute_87','compute_89','compute_90',
'compute_90a','lto_50','lto_52','lto_53','lto_60','lto_61','lto_62','lto_70',
'lto_72','lto_75','lto_80','lto_86','lto_87','lto_89','lto_90','lto_90a',
'native','sm_50','sm_52','sm_53','sm_60','sm_61','sm_62','sm_70','sm_72',
'sm_75','sm_80','sm_86','sm_87','sm_89','sm_90','sm_90a'.

torch-points-kernels setup file explicitly adds support for the architecture sm_35 for a currently unknown reason (see more context).

This PR removes this additional build flag.

Also, I refactored unit tests workflow to adjust some stuff to our current setup.

The plan is to issue a new release, 0.7.2 after merging this.

[griegler]

curious, why are the other dependencies not pinned to a specific version?

[bugra-yilmaz]

I've received an error from Torch - Numpy interaction and this is the version we have in OSS dependencies currently.

The public repository doesn't have any changes to the unit tests workflow for 2 years - they don't support Python 3.8 yet so I thought this is already much better than what there was 😛

[nora-promaton]

How did this work with the matrix, did we test for all os x python version x torch versions? And now we only will test with ubuntu latest and torch?

[nora-promaton]

This change was not listed as a goal of this PR. Why don't we need macos and windows anymore?

[bugra-yilmaz]

We don't run anything on macOS or Windows so I didn't want us to waste time / resources with those.