Fix and simplify pytorch backend: array, einsum, from_numpy

[luisfpereira]

This PR acts specially in array for torch backend. It avoids unnecessary conversions to numpy.

Additionally, dtype has been added to from_numpy to cast when required. It needs to be extended to the other backends for compatibility, but it will play a role when scipy is used with torch or tensorflow (to ensure precision is kept).

einsum is also simplified. @ninamiolane the tests are not failing with this simplification, but I'm not sure if I'm missing something (I was not able to understand why einsum had such a complex wrapper). If you have an example that justifies the wrapper, please let me know so I revert the changes there.

It closes #1537 and starts paving the path to better performance in GPU.

[codecov]

Codecov Report

Merging #1550 (7324258) into master (f284cc3) will decrease coverage by 0.77%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1550      +/-   ##
==========================================
- Coverage   90.73%   89.96%   -0.76%     
==========================================
  Files         102      102              
  Lines       10414    10288     -126     
==========================================
- Hits         9448     9255     -193     
- Misses        966     1033      +67     

Flag	Coverage Δ
autograd	?
numpy	87.15% <ø> (?)
pytorch	80.97% <100.00%> (-0.08%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
geomstats/_backend/pytorch/__init__.py	95.60% <100.00%> (+0.40%)	⬆️
geomstats/information_geometry/gamma.py	55.04% <0.00%> (-35.90%)	⬇️
geomstats/geometry/hypersphere.py	80.91% <0.00%> (-7.11%)	⬇️
geomstats/learning/geodesic_regression.py	84.42% <0.00%> (-5.19%)	⬇️
geomstats/geometry/grassmannian.py	93.69% <0.00%> (-1.05%)	⬇️
geomstats/learning/frechet_mean.py	96.00% <0.00%> (-0.88%)	⬇️
geomstats/geometry/special_euclidean.py	88.68% <0.00%> (-0.24%)	⬇️
geomstats/_backend/__init__.py	79.37% <0.00%> (ø)
geomstats/_backend/autograd/linalg.py
geomstats/_backend/autograd/__init__.py
... and 11 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f284cc3...7324258. Read the comment docs.

[ninamiolane]

Thank you for this great simplification! I can't remember exactly why einsum was this complicated, it could be to allow einsums:

• of the type ("ijk, jkl->ijl", a, b) i.e. with tensors of order 3 and more
• of the type ("ijk, jkl, mjkl,->mijl", a, b, c) i.e. with 3 tensors or more
• with ellipsis, i.e. of the type ("...ijk, ...jkl..., mjkl,->...mijl...", a, b, c)
• a combination of the above,
  and it could be that previous versions of pytorch were not allowing it but new ones do?

In any case, if tests do not fail we can keep it like this... and remember where to look for if we encounter problems.

Thanks!