Minor fixes for PARAFAC2 (#263 and more)

[MarieRoald]

This pull request fixes some small issues with the PARAFAC2 code

• Changed to a relative loss decrease stopping condition
• Added an absolute loss decrease stopping condition which defaults to 1000 times the machine precision (necessary for convergence on noise-free PARAFAC2 tensors) (discussion here: The stopping condition of PARAFAC2 uses an absolute loss decrease condition, not a relative loss decrease condition #263 )
• Changed the default maximum number of iterations to 2000
• Added nn_modes to the docstring (thanks @IsabellLehmann 😃)
• Updated the Parafac2 class to have the same arguments and docstring as the parafac2
• Added tests for stopping conditions of PARAFAC2

The added absolute tolerance and new maximum number of iterations were also based on the MATLAB code by Rasmus Bro (available here: http://www.models.life.ku.dk/algorithms).

@JeanKossaifi, before we can merge this, we should update the version number in the documentation (see the todos on line 185, 203, 378 and 396 of the _parafac2.py file). What would be the right version number?

Also, when I built the sphinx documentation, I noticed that there were no special formatting for the versionchanged HTML class. Maybe we should update the CSS?

[codecov]

Codecov Report

Merging #267 (fee5fd2) into main (23ba632) will increase coverage by 0.02%.
The diff coverage is 52.00%.

@@            Coverage Diff             @@
##             main     #267      +/-   ##
==========================================
+ Coverage   87.26%   87.28%   +0.02%     
==========================================
  Files          91       91              
  Lines        4553     4563      +10     
==========================================
+ Hits         3973     3983      +10     
  Misses        580      580              

Impacted Files	Coverage Δ
tensorly/decomposition/_parafac2.py	80.13% <25.00%> (-2.14%)	⬇️
tensorly/decomposition/tests/test_parafac2.py	100.00% <100.00%> (ø)
tensorly/decomposition/_cp.py	78.35% <0.00%> (-0.38%)	⬇️
tensorly/contrib/decomposition/_tt_cross.py	92.45% <0.00%> (+3.14%)	⬆️

---

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 23ba632...fee5fd2. Read the comment docs.

[JeanKossaifi]

Awesome, thanks @MarieRoald!

I've bumped the current version to '0.6.1'. This is a good point, we can add it to the scss and extend either the generic admonition or a specific one, e.g. note:

https://github.com/tensorly/tensorly-sphinx-theme/blob/d7d7a6a1c97535481296015b1778912e74760053/src/sass/tensorly_style.scss#L187

[MarieRoald]

I added the version number, and I also changed the absolute tolerance to 1e-13, since 1000*eps was too large for the single-precision backends.

[JeanKossaifi]

Thank you @MarieRoald, merging!