[SYSTEMDS-3636] Improved ultra-sparse tsmm left w/ sparse output

[christinadionysio]

This patch provides the support for left transposed ultra-sparse tsmm.
Similar to the the implementation of the right transpose ultra-sparse tsmm,
binary search is used to populate the upper triangular part of a sparse output matrix.

[Baunsgaard]

how much faster is it?

[christinadionysio]

how much faster is it?

@Baunsgaard I only tested the t(X) %*% X operation on ultra sparse matrices locally for now. Using the germany_osm dataset we achieve a runtime of ~12s after the optimisation. (It is infeasible to run it locally without the optimisation, since a dense matrix is allocated and therefore a Java heap space exception occurs.)
When I run the operation on an ultra sparse matrix with 30000 rows and cols I get a speedup of 17x or 30x depending on which of the two methods is used.

[Baunsgaard]

a few comments for improvements, hope you can decrypt it.

[christinadionysio]

After the changes the runtime of the matrixMultTransposeSelfUltraSparse method improved by 2.5x.

[Baunsgaard]

LGTM

[Baunsgaard]

Thanks for the PR, it is now merged!