TST make sure test_pca_sparse passes on all random seeds

[jeremiedbb]

Fixes #28857

The array that we're comparing have a very wide range of values, from 1e-8 to 1e0. Having a same rtol for all is kind of problematic. In this PR I introduced an additional atol, useful for the very low values. In the plot below I show the absolute difference of the components vs the absolute value of the components to illustrate the need of an atol:

We see that for very small values of the components, the absolute diff doesn't foolow the same trend as for larger values. It looks like a plateau.

The following plot shows the relative diff of the components vs the absolute values of the components. It's clearly not constant or even upper bounded (if we were to extrapolate for even smaller values). This goes against what we assume when we write assert_allclose(X1, X2, rtol=constant).

This is symptomatic of a general issue that we have through the whole project for comparing arrays element-wise. I made a quick fix here to make the CI green, but I think that it should be improved in general (I've been thinking about that for a while: it's not an easy problem and I haven't found a satisfying solution yet).

[github-actions]

✔️ Linting Passed

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

_{Generated for commit: 928f723. Link to the linter CI: here}

[jeremiedbb]

There's some randomness that I can't explain (and can't reproduce locally). The failing tests are not exactly the same in b2d30f9 and in e0448a3. The failing tests from #28857 isn't always failing either.

[ogrisel]

Maybe we could check that (pca.explained_variance_ > np.finfo(X.dtype).eps).all() before running any other assertion. If n_components is large enough and density is small enough, it's possible that the matrix rank of the data is lower than n_components and therefore, some components will be random.

[jeremiedbb]

I checked and they're always a lot larger. We're only looking for a max of 10 components. Even with the min density of 0.01, all 10 components always have the same approx explained variance of 0.4%, for all seeds.

The kind of randomness I mention here has a very very small impact. It changes the final digit once in while. I think we can ignore it for now, the new tols are robust to that.

[ogrisel]

LGTM then. Thanks for the PR.