Remove per-launch heap allocation in CPU __run

[simone-silvestri]

the current __run function leads to heap allocation because it boxes the args.
Apparently the problem is with len and rem that are boxed because of being defined inside an if condition which leads to boxing of the whole args. In Oceananigans, these args are typically very large (we can get easily to MB) which in turn leads to a lot of GC getting called.

MWE:

julia> ci = only(code_lowered(KernelAbstractions.__run, NTuple{6, Any}));

julia> !any(e -> occursin("Box", string(e)), ci.code)
false

This is a backport given that KA main switched to POCL,

cc @giordano

[giordano]

@simone-silvestri you need to merge on top of the remote target branch to get the buildkite pipeline work (that was fixed by #717)

[simone-silvestri]

Thanks @vchuravy, can we release a version 0.9.42?

edit: Ah already done! Thanks