Support large sparse matrices in SGD* and SequentialDataset

[jnothman]

Support for sparse matrices having indices, indptr, row or col attributes with int64 dtype was recently added or confirmed for most sparse-supporting estimators.

SGDClassifier and SGDRegressor do not yet support such large sparse matrices (they use accept_large_sparse=False). We could try to fix this.

[TomDLT]

Supporting int64 indices in sparse SGD mainly resorts to updating CSRDataset.
As Cython fused types do not work with class attributes, this will need to use the same template workaround as in #11155, which might become quite heavy.
I wonder if there is not a cleaner way based on fused types, maybe by dropping SequentialDataset.

[rth]

As Cython fused types do not work with class attributes, this will need to use the same template workaround as in #11155, which might become quite heavy.

That or some void pointer arithmetics maybe (cf http://blog.yclin.me/deep/learning/2016/08/08/Fused-Types-Limitation/). Neither seems ideal.

Also actually there is an existing PR for this issue in #6889 . This was also previously reported in #5776.