I'd like to add a .sparse accessor to DataFrame, to assist with deprecating SparseDataFrame.

It'll contain

• from_spmatrix (part of the SparseDataFrame constructor)
• to_dense (SparseDataFrame.to_dense)
• to_coo (SparseDataFrame.to_coo)
• density

A few design questions:

1. When should the _validate raise?
   a. When there are no sparse columns
   b. When there is any non-sparse columns
   c. Never.

It's slightly easier to implement if we assume everything is sparse.

2. Return value of DataFrame.sparse.density. If we mirror SparseDataFrame.density, this returns a float. Would it be more useful to return a Series with the density of each column? (and users can .mean() if they want the average density)

I believe that with these methods, the essentially all the functionality of SparseDataFrame will be replicable with a DataFrame of sparse values (the main exception being an expanding __setitem__ creating a sparse column by default; but that's OK to not provide that functionality).