revise the handling of backend and handle properties in Program

[ksimpson-work]

Program is either backed by nvrtc (cuda-bindings), or the Linker object (cuda-core abstraction).

Come up with, and implement a scheme for the Program.handle and Program.backend slots whose return values are currently ill defined and vary depending on the Program's backend.

see thread #317 (comment)

[ksimpson-work]

We should talk about this in the thursday cuda.core meeting. Specifically the implications to #370 and #192 are potentially noteworthy

[leofang]

btw in the original design, .handle/.from_handle() are meant for C-Python interoperability, where we can take a handle from cuda.core and pass it to a foreign C library, or vice versa. It seems with #370 we might start making the story more complicated than necessary. I need to think about it while working on #180/#370.

[ksimpson-work]

copy that