EFF Optimize memory usage for sparse matrices in LLE (Hessian, Modified and LTSA)

[giorgioangel]

What does this implement/fix? Explain your changes.

This PR optimizes memory management with sparse matrices when using Modified Locally Linear Embedding.

Before this PR, a numpy NxN array was created, filled, and then converted to sparse. The creation of the said array can require huge memory when dealing with a large dataset.
On the dataset I was working with, the algorithm tried to allocate 400GB of RAM lol...

In the current PR, when M_sparse is true, the algorithm creates directly a sparse matrix, greatly reducing the memory requirements.

[github-actions]

✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit: c2e82b3. Link to the linter CI: here}

[glemaitre]

Could you add an entry in the changelog doc/whats_new/v1.5.rst to acknowledge the change.
Could you also have the same approach for the other method to be consistent. We should also make sure that we test all the combination in the tests.

[glemaitre]

I am quite worry that with a bug, we did not have any test failing. We should make sure to have something minimal here.

[glemaitre]

Sorry @giorgioangel I did not have time to follow-up before the release. I'll add the milestone for 1.6 and make a review. I'll sort out any conflict and ping someone else for a second review.

[glemaitre]

LGTM. Since we have a test that check for all solvers and all methods, then it means that we did not introduced a regression. So I would advocate that we don't need any additional tests

[OmarManzoor]

LGTM. Thanks @giorgioangel