[FEA] Make block, thread, and warp indices unsigned

[gmarkall]

Block, thread, and warp indices would ideally be unsigned, as in CUDA C/C++. This would:

• Reduce computation and register usage (due to handling signs)
• Align better with CUDA C/C++

However, due to Numba's typing it can result in float indices being generated rather than int ones (e.g unsigned + signed = float or something like that), or 32-bit values becoming 64-bit ones (as observed in numba/numba#6112 (comment))

This was started in numba/numba#6112 - completion of this PR would be sufficient to implement this feature request, but it may not be as simple as getting the test suite to behave identically - effects may be observed in more complex programs where types change.

An alternative path may be to define a separate type for thread indices that is more resistant to upcasting when used in computations, but the exact solution is unclear.

[gmarkall]

I just started going over the original PR and noticed that:

• It already implements a special type for indices (tid)
• The tid type can be promoted to 64 bits:
  • Via tcr.promote_unsafe(types.tid, types.int64) in https://github.com/numba/numba/pull/6112/files#diff-db8d853bfa25e99ecb857628481643fa8915b179607ce7db7d29df01520c454d

Perhaps the solution is to remove the possible promotion of the tid type to 64 bits, and would be the next thing to investigate.

[gmarkall]

Draft PR that ports the changes over to Numba-cuda is in #110.