[NativeCPU] Simplify enqueue.

[hvdijk]

We were creating excessive numbers of threads. When we know we want a given amount of threads, just divide the number of workgroups by the number of threads and have each thread process that many workgroups.

This implementation also means we no longer need to resize workgroups, which was not generally safe.

[uwedolinsky]

It seems this PR also removes the unsafe enqueuing optimization for nd_range kernels. If so it's probably worth also mentioning this in the description.

[hvdijk]

It seems this PR also removes the unsafe enqueuing optimization for nd_range kernels. If so it's probably worth also mentioning this in the description.

Done, thanks.

[hvdijk]

@intel/llvm-gatekeepers This can be merged, thanks.