Describe the workflow you want to enable

import scipy.sparse as ss
from sklearn.decomposition import TruncatedSVD

X = ss.random(1000, 100, format='csc')
tsvd = TruncatedSVD().fit(X)

tsvd.transform(X)

Currently, in the last line, transform(X) will first convert X from csc to csr and then apply a dot product. I think this conversion is unnecessary, costs additional time, and one should better rely on scipy.sparse (or safe_sparse_dot) to convert if necessary.

Describe your proposed solution

Allow csc in

Additional context

A PR with a benchmark result might be nice.