[WIP] Faster backend compatible ot.dist

[rflamary]

Types of changes

The current implementation of ot.dist uses the cdist foncton form sipy on non euclidean costs which makes it very slow. This PR is an attempt at making it faster by implementing the most common distances using the backend.

For the moeme,t backend computation are ot always faster

for instanec on a 1024x200 matrix X we have
n= 1024 , d= 200

	cdist (scipy)	dist 0.9.5	ot.dist (numpy)	dist (torch)	dist (torch, ten)	dist (torch,GPU)
euclidean	0.942832	0.0836014	0.0897373	0.0932314	0.00872684	0.429742
sqeuclidean	0.798189	0.0778745	0.0659651	0.00911853	0.0744656	0.00241347
cityblock	1.2758	1.26534	6.4717	0.42605	2.14289	0.0770559
cosine	0.971009	0.970279	0.473443	0.0740105	0.00881105	0.00222607
minkowski	0.847159	0.843644	6.97861	0.614996	3.37625	0.0993479
correlation	1.01531	1.01237	0.0785626	0.0948708	0.012194	0.00431579

Motivation and context / Related issue

How has this been tested (if it applies)

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR and Issue fix to the RELEASES.md file.

[codecov]

Codecov Report

Attention: Patch coverage is 85.71429% with 10 lines in your changes missing coverage. Please review.

Project coverage is 97.02%. Comparing base (1761d0b) to head (3be92a4).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #701      +/-   ##
==========================================
- Coverage   97.05%   97.02%   -0.03%     
==========================================
  Files          98       98              
  Lines       19955    20008      +53     
==========================================
+ Hits        19367    19413      +46     
- Misses        588      595       +7