SVD limitations

[javicid]

The current implementation of SVD has the following limitations:

• Convergence checking needs to synchronize data (could be improved with exceptions mechanism)
• Convergence criteria should be revised because it was not completely clear to me how to compute it
• n_rows needs to be greater than n_cols (this could be solved by implementing a row-wise algorithm or by transposing the ds-array)
• Uses the simplest pairings for columns, which might not be the best for parallelism (other pairings should be explored)
• Doesn't work for irregular ds-arrays
• It might be better to compute V.T instead of V
• The input array needs at least 2 column blocks (we could use numpy directly or a row-wise algorithm for ds-arrays with one column block)