CPU and GPU operator customization based on Taichi

[chaoming0625]

Please:

• Check for duplicate requests.
• Describe your goal, and if possible provide a code snippet with a motivating example.

See taichi-dev/taichi#6367

[Routhleck]

Expose the import_cpu_memory and import_cuda_memory C API so that taichi's NdArray can directly use the memory address allocated by XLA to eliminate the memory copy consumption. taichi-dev/taichi#8368

[chaoming0625]

Does the CUDA backend also work now?

[Routhleck]

YES, I just tested it on CUDA backend and the results are correct!

[chaoming0625]

See #527

[chaoming0625]

See #553