Multikernel option for bdsqr optimization

[tfalders]

As previously discussed, I have been experimenting with optimizing BDSQR by using multiple kernel launches, with device synchronizations to determine the iterative loop's stopping condition.

Broadly speaking, I have made the following changes:

• bdsqr_t2bQRstep and bdsqr_b2tQRstep have been combined into a single device function. This was intended to allow different split blocks to be processed in the same thread group, but it wasn't performant. Still, I've kept this change.
• bdsqr_kernel has been divided into multiple different kernels, with the iterative loop moved to the CPU.
  • bdsqr_compute determines the shift for the current split block and applies the QR step to D and E. If nv, nu, and nc are less than the specified switch size, it will also update the singular vectors.
  • bdsqr_rotate updates the singular vectors using multiple thread groups.
  • bdsqr_update_endpoints updates the endpoints of each split block, and spawns new split blocks if zeroes are found in the middle of a split block.
• bdsqr_chk_completed is a new kernel that will determine if the stopping criterion has been met for each batch instance.
• The work and splits_map arrays have been expanded to accommodate new information, and a new completed array has been added to hold information about problem status. See the comment block on line 316 for an explanation of how data is stored in these arrays.