Move callback wrappers to the Python layer

[vzhurba01]

Sounds like a good idea to

• move the lifetime-related changes to another PR
• ensure both driver/runtime Python layers have the lifetime management for callbacks
• ensure both driver/runtime Cython layers have no lifetime management for callbacks (as if the CUDA C APIs are called from Cython, which was the intent AFAIK)

Originally posted by @leofang in https://github.com/NVIDIA/cuda-python/pull/517#discussion_r2010799701