Dense sketching: only generate the submatrix needed for a requested multiplication

[rileyjmurray]

This is a follow-up to PR #51 and will resolve issue #36.

The changes here warrant an update to the web documentation. That'll be handled in separate PR.

Changes to functions for generating dense sketching operators

• Renamed fill_rsubmat_omp to fill_dense_submat_impl.
• Defined a new function, fill_dense_submat, that wraps fill_dense_submat_impl.
  • This new function handles row-major or column-major data.
  • This new function dispatches the OP template parameter in fill_dense_submat_impl.
• Made several modifications to fill_buff
  • It's now a one-line wrapper of fill_dense_submat.
  • Its name has been changed to fill_dense.
  • The positions of its first and second arguments have been swapped.
• Defined an overload of fill_dense that only accepts a sketching operator (so as to match the semantics of fill_sparse).
• Removed fill_rmat, which was only used in the old fill_buff implementation.

Changes to functions for applying dense sketching operators

LSKGE3 and RSKGE3 were changed so that if the input sketching operator S0 satisfied the condition !S0.buff, then we would

1. define a new sketching operator S with the same shape as $\mathrm{submat}(\texttt{S0})$ and a BlackBox distribution family, then
2. populate the buffer of S on an appropriate submatrix of S0 (possibly all of S0), before
3. making a recursive call to LSKGE3/RSKGE3 with (S0, i_os, j_os) replaced by (S, 0, 0).