GPU - gen fallback to shared if error

[jeremylt]

Ok, there's two main reasons why a gen operator might fail

1. too much shared memory needed to compile

2. too many resources needed to run kernel

This PR adds a mechanism to automatically detect these issues and fall back to a shared operator.

[jeremylt]

@zatkins-dev can I get your thoughts on the overall approach? This should basically let people use /gpu/[cuda,hip] and all unsupported operators will automatically fallback

[zatkins-dev]

Yeah, I like the approach. I definitely think that it's good for users to be able to use /gpu/[cuda|hip] without worrying about support for whatever their use case is on a particular implementation

[zatkins-dev]

Overall I think this looks good, with a few minor things