CUDA: decouple LSPAN from BLOCKSIZE in leg_m_kernel() for greater flexibility and better performance with multiple warps per block.

HIP/CUDA: new flag SHT_ALLOW_SH2ISH_FUSE to fuse sh2shioka() into leg_m_kernel() for better performance.

CUDA: allow NW>1 for HI_LLIM in leg_m_kernel() with polar optim, leading to performance improvements (5-10% on V100)

CUDA: in (i)leg_m_kernel(), swap z and y dimensions of grid (better cache usage?), for 3% performance improvement on V100

fix(batch gpu vector transforms): in batch mode, the on-device vector transform functions cu_SH*_to_spat() and cu_spat_to_SH*() were wrong.

feat(batch): shtns_set_batch() must now be called before shtns_set_grid*(). Avoids planning for 1 transform before many.

feat(cuda): change the way cuda streams are managed; cushtns_set_streams() should now be called just after shtns_create()