Support fetching for particular CUDA capabilities

[sayakpaul]

To reduce download costs.

[awneesht]

I hit the adjacent diagnostic problem while testing Hub kernels on
older hardware, and wanted to check whether it belongs here or
elsewhere before I put work into it.

The failure I keep seeing isn't resolution.get_kernel returns a
build fine, it's that the build loads and then fails at launch,
because my compute capability wasn't among the compiled targets and
there's no PTX to JIT from. From the caller's side that's an opaque
error some distance from the cause.

What I've been doing by hand: cuobjdump --list-elf on the resolved
.so, compare against torch.cuda.get_device_capability(), then run
the kernel's correctness check and time it against eager.

Two questions:

1. Is surfacing "your device isn't a compiled target for this
   variant" in scope for this issue, or is More (optional) load time verification #715's load-time
   verification the right home for it?
2. If it's wanted, would you prefer it inside kernels, or as a
   separate diagnostic that reads the resolved artifact?

Happy to prototype whichever fits. Mainly want to avoid building
something parallel to what's already planned.