Basic cl_khr_subgroups implementation for CPU

The subgroup is always the X dimension, and there is only one
subgroup in flight at the same time. Thus there are Z*Y SGs
per WG. It passes the CTS test even though independent forward
progress is not supported because there's only one SG in flight
at the same time making progress.

Also preliminary shuffle and ballot implementations which work
only with uniform execution for now.

The cross-WI data exchange is implemented via
__pocl_{wg,local_mem}_alloca(), an internal function that
allocates "local memory" (thread stack) dynamically.