Make block, thread, and warp indices unsigned

[gmarkall]

This ports numba/numba#6112 to numba-cuda, as outlined in #109.

Note that for this patch, we don't change the type of grid() and gridsize() because these need to be 64 bit (as discovered in numba/numba#9229 and fixed in numba/numba#9235).

We need to patch the as_dtype() function, which is a little unfortunate, but there's no API for extending its behaviour at present.

Current status: in progress, as items from the original PR need to be addressed (as per this comment):

• Audit whether type widening still occurs
• Examine the cases of increased register usage (which is likely due to policies in the PTX -> SASS compiler rather than anything changed in numba-cuda).

We still allow unsafe promotion of tid to 64 bits, because without it, comparisons with other 64-bit operands cannot occur.