Types of changes

This PR aims at

• implementing the detach function in the backend to allow speedup on CPU/GPU in some solvers (which was already done in a previous PR but with limited doc).
• Implement variants of Sinkhorn where computations are detached and gradients at convergence is returned instead

This PR should solve #565 and greatly limit memory for sinkhorn when computing gradienst wrt the value.

In order to use implicit diffeerntiation one needs to set the grad parameter in ot.solveand ot.solve_sampleas such

sol = ot.solve(M, a, b, reg=10, grad='implicit')
sol.value.backward()
# beware with  grad='implicit', sol.value_linear and sol.plan are not differentiable (not implemented yet).

On a simple example with pytorch arrays with required gradients, I has a 1000x gain in memory for solving the problem when a large number of sinkhorn operations are needed.

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.