Change `gpu_stump` to Pearson Correlation

[seanlaw]

trafficstars

Currently, gpu_stump uses the sliding dot product approach to computing matrix profiles rather than Pearson Correlation. In theory, we should be process each row at a time (rather than traversing diagonals) and instead of storing QT from the last row, we store the covariance from the previous row.

Doing this should result in a nice speed improvement and makes the codebase more consistent with the stump functions since they also use Pearson Correlation