Make it possible to use ARPACK or randomized_svd solvers in PLSSVD

[ogrisel]

As initially discussed in #18302 (comment) it might be interesting to add an extra constructor param to PLSSVD to select ARPACK or the sklearn's randomized_svd solver instead of the default LAPACK solver (from scipy.linalg.svd).

But the ARPACK and randomized_svd are non-deterministic so we would also need to add a random_state parameter.

Careful benchmarking to evaluate the speed vs numerical or statistical accuracy trade-off should be conducted to:

• help the user choose the value of this parameter (both in the docstring and the user guide)
• suggest an "auto" strategy to automatically select a good solver based on the shape of the data and the n_components parameter, similar to what is done in the PCA and TruncateSVD estimators. This "auto" parameter shall become the default after the usual deprecation period.

[ogrisel]

This is related to #12069 that does a similar thing for KernelPCA.