Calling fit() or fit_predict() with a csr matrix as input crashes the code (when affinity != 'precomputed').

I've investigated the error a bit and found that X = np.asarray(X) in the fit method of affinity_propagation_.py is causing the error and should be replaced by sth like X = check_array(X, accept_sparse='csr') as is done for DBSCAN.

Note:
Just re-checked, the problem does not happen with the latest stable release (scikit-learn==0.15.2) as the problematic line X = np.asarray(X) is not in the fit() method.

Minimal example to reproduce the error:

from sklearn.cluster import AffinityPropagation
from sklearn.datasets import fetch_20newsgroups_vectorized

b = fetch_20newsgroups_vectorized()
affinity = AffinityPropagation()
affinity.fit(b.data)

Setup:
scikit-learn==0.16.dev0