Improve dtype handling.

[tpn]

This PR improves dtype handling for cuda.cooperative routines, allowing a much wider range of values to be passed to any of the Pythonic CUB routines that accept a dtype param. Per the normalize_dtype_param() docstring, the logic is as follows:

    - If the dtype is already a numba type, return it as is.
    - If the dtype is a valid numpy dtype, convert it to the corresponding
      numba type.  Note that this applies to both `np.int32` and
      `np.dtype(np.int32)`.
    - If the dtype is a string:
        - If there's a period in the string, ensure it's because the string
          starts with "np." and is followed by a valid numpy dtype.  Otherwise,
          raise a ValueError in both cases: if there's a period belonging to
          something other than the leading "np.", or if the following numpy
          type isn't valid.
        - If there's no period, assume the type is referring to a numba type.
          If not, raise a ValueError.  If it is, return the corresponding numba
          type.

This is for #3914.

[github-actions]

🟩 CI finished in 1h 00m: Pass: 100%/1 | Total: 1h 00m | Avg: 1h 00m | Max: 1h 00m

• 🟩 python: Pass: 100%/1 | Total: 1h 00m | Avg: 1h 00m | Max: 1h 00m

  🟩 cpu
    🟩 amd64              Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  🟩 ctk
    🟩 12.8               Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  🟩 cudacxx
    🟩 nvcc12.8           Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  🟩 cudacxx_family
    🟩 nvcc               Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  🟩 cxx
    🟩 GCC13              Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  🟩 cxx_family
    🟩 GCC                Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  🟩 gpu
    🟩 rtx2080            Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  🟩 jobs
    🟩 Test               Pass: 100%/1   | Total:  1h 00m | Avg:  1h 00m | Max:  1h 00m
  

👃 Inspect Changes

Modifications in project?

	Project
	CCCL Infrastructure
	libcu++
	CUB
	Thrust
	CUDA Experimental
+/-	python
	CCCL C Parallel Library
	Catch2Helper

Modifications in project or dependencies?

	Project
	CCCL Infrastructure
	libcu++
	CUB
	Thrust
	CUDA Experimental
+/-	python
	CCCL C Parallel Library
	Catch2Helper

🏃‍ Runner counts (total jobs: 1)

#	Runner
1	linux-amd64-gpu-rtx2080-latest-1

[brycelelbach]

Looks good except for lowering int->int32 and float->float32

[github-actions]

🟩 CI finished in 58m 48s: Pass: 100%/1 | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s

• 🟩 python: Pass: 100%/1 | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s

  🟩 cpu
    🟩 amd64              Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  🟩 ctk
    🟩 12.8               Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  🟩 cudacxx
    🟩 nvcc12.8           Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  🟩 cudacxx_family
    🟩 nvcc               Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  🟩 cxx
    🟩 GCC13              Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  🟩 cxx_family
    🟩 GCC                Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  🟩 gpu
    🟩 rtx2080            Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  🟩 jobs
    🟩 Test               Pass: 100%/1   | Total: 58m 48s | Avg: 58m 48s | Max: 58m 48s
  

👃 Inspect Changes

Modifications in project?

	Project
	CCCL Infrastructure
	libcu++
	CUB
	Thrust
	CUDA Experimental
+/-	python
	CCCL C Parallel Library
	Catch2Helper

Modifications in project or dependencies?

	Project
	CCCL Infrastructure
	libcu++
	CUB
	Thrust
	CUDA Experimental
+/-	python
	CCCL C Parallel Library
	Catch2Helper

🏃‍ Runner counts (total jobs: 1)

#	Runner
1	linux-amd64-gpu-rtx2080-latest-1