Add lowerkernel CPU backend

[wsmoses]

For debugging ln a cuda less machine having a cpu backend to the kernels call op would be useful

We can basically just make a new funx with outermost parallel fors for block and thread, then similarly jit compile it and insert into a custom call expr

For now that should suffice if there's no shmem or sync, we can do a polygeist style lowering after GB

[vchuravy]

The better way is likely to go to SPIR-V and OpenCL+POCL. A COU backend as you describe is still tricky since you need to forward variables between the loops.

[wsmoses]

without sync/shmem there's only one loop nest, so there shoulod be no need to forward any variables