Introduce primal_cost and dual_cost for Sinkhorn outputs (only primal for LR) for arbitrary geometries.

[marcocuturi]

• Solves Define a ot_cost property for sinkhorn and sinkhorn_lr outputs #180 by introducing a property to compute primal_cost, the bare cost of transportation (i.e. the dot product between a coupling matrix and the corresponding cost matrix) for sinkhorn and sinkhorn_lr (it was already there for the latter, just renamed). While trivial for standard geometries, notably generic Geometry objects, this is a bit more complicated for PointCloud and required the ability to cast an arbitrary Grid (itself with possibly low rank costs at each dimension) into a low-rank geometry.
• Because casting Grid to LRCGeometry requires casting all cost matrices into LRC, introduce the possibility, usingrank=0 or rank>=min(n,m) to default to SVD to cast an exact, low-rank factorization.
• Introduce dual_cost (much easier to compute) for outputs of sinkhorn and sinkhorn_lr.
• While testing for primal_cost in Grid, exhibit bug Grid geometry handling of 0 weights seems buggy #185.
• Correct stopping criterion in balanced sinkhorn to amend for the fact that in the case updates are parallel, both marginals should be computed.
• As a result, modify the computation of symmetric terms in sinkhorn divergence computations as using parallel updates by default, and test this leads to improved or equivalent computational times.

[codecov-commenter]

Codecov Report

Merging #184 (1f2104a) into main (156ef2b) will decrease coverage by 0.01%.
The diff coverage is 97.24%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #184      +/-   ##
==========================================
- Coverage   89.28%   89.26%   -0.02%     
==========================================
  Files          51       51              
  Lines        5187     5234      +47     
  Branches      527      529       +2     
==========================================
+ Hits         4631     4672      +41     
- Misses        430      434       +4     
- Partials      126      128       +2     

Impacted Files	Coverage Δ
ott/tools/sinkhorn_divergence.py	98.24% <ø> (+3.60%)	⬆️
ott/geometry/low_rank.py	96.42% <66.66%> (-0.66%)	⬇️
ott/solvers/linear/sinkhorn_lr.py	96.25% <83.33%> (-0.34%)	⬇️
ott/geometry/geometry.py	92.15% <97.29%> (-0.10%)	⬇️
ott/geometry/graph.py	94.40% <100.00%> (+0.03%)	⬆️
ott/geometry/grid.py	86.66% <100.00%> (-4.02%)	⬇️
ott/geometry/pointcloud.py	85.54% <100.00%> (+0.25%)	⬆️
ott/solvers/linear/sinkhorn.py	97.85% <100.00%> (-0.29%)	⬇️
ott/solvers/quadratic/gromov_wasserstein.py	93.05% <100.00%> (ø)
... and 3 more

[marcocuturi]

Changed my mind, ot_cost was a bit ambiguous and clearly overloaded. I prefer primal_cost (and dual_cost), which are (a bit) more explicit. primal_cost will always be the linear term of a linear OT problem (or, by extension, in a future PR, the quadratic/fused for a quadratic/fused pbm).