[cudaaligner] myers_gpu uses workspace per block now

[ahehn-nv]

• aligner_global_banded_myers launches thread-blocks depending on the SM count of the GPU.
• Each thread-block has its own workspace and processes multiple alignment tasks sequentially.
• There are two different workspace sizes: n large workspaces for the n largest alignments, many "small" workspaces.

This allows to compute more alignments in a batch due to reduced memory usage per alignment.