TaskPolicy<Cuda> performance requires teams mapped to warps

[hcedwar]

The sparse cholesky example using TaskPolicy has 12% occupancy.
Probable fix is to support mapping of thread teams onto warps where the block size is multiple warps.
This requires enhancement of the thread team's team-level collectives and nested parallel operations.
A possible redesign is to replace the team-size integer parameter with a "large/small" value that allows the TeamPolicy and TaskPolicy to choose a device-appropriate value for large and small.