Enable multiclass svm for sparse input

[mfoerste4]

This commit enables multiclass SVM for sparse input. Previously this was deactivated as the input_to_host_array functionality does not support sparse arrays, but the data has to be piped through sklearn classes which requires host data.

@dantegd , this is a local workaround to enable a input_to_host_array for sparse data without the complexity of providing the whole functionality of that function. Please have a look whether this is an acceptable solution for this use case.

FYI, @tfeher

[tfeher]

Thanks Malte for the PR, it looks good overall. Just one question about testing: do we have only the pickle test for this feature? If yes, could we improve coverage in test_svm.py?

[mfoerste4]

Thanks Malte for the PR, it looks good overall. Just one question about testing: do we have only the pickle test for this feature? If yes, could we improve coverage in test_svm.py?

@tfeher , I extended an existing multiclass test to also use sparse inputs.

[csadorf]

Just one request, otherwise looks good. Make sure to update the base branch to branch-23.12!

[tfeher]

Thanks @mfoerste4 for adding the test! Apart from the last open question about the helper placement the PR looks good to me!

[mfoerste4]

Just one request, otherwise looks good. Make sure to update the base branch to branch-23.12!

@csadorf, I believe the failing CI checks are unrelated to this PR. Could you have a look at the changes requested?

[csadorf]

LGTM!

[csadorf]

/merge

[csadorf]

@mfoerste4 @tfeher I think the cudaErrorMisalignedAddress:misaligned address errors we are seeing in the wheel tests are actually not a fluke, but related to the change set in this PR. I don't have a hypothesis yet as to what might be causing it.

[csadorf]

@mfoerste4 Thanks for the fix. Tests are currently expected to fail because of an upstream incompatibility between raft and rmm.

[mfoerste4]

@mfoerste4 Thanks for the fix. Tests are currently expected to fail because of an upstream incompatibility between raft and rmm.

@csadorf , I was not able to reproduce the issue locally, I just hope it is a duplicate of the bug that i came across yesterday.