Support on-disk caching in the CUDA target - #8089

This primarily uses Numba's existing infrastructure for on-disk,
caching, and only requires filling in / connecting up some small extra
parts in the CUDA target. These changes are:

- Fix `_reduce_states()` and `_rebuild()` ion `CUDACodeLibrary`. These
  were never used before, so some simple logic errors persisted.
- Add a `magic_tuple()` function to the `JITCUDACodegen`. This is used
  to uniquely describe the codegen behaviour, which for CUDA is runtime
  version- and compute capability-specific.
- Wire up the `cache` kwarg in the `@cuda.jit` decorator to enable
  caching.
- Add some properties to CUDA `_Kernel` objects that the caching
  implementation expects to find, and modify its `_reduce_states()` and
  `_rebuild()` methods to store the signature instead of the list of
  argument types (which aligns it more closely with the design of a
  `CompileResult`).
- Implement a `CUDACache` class for the core caching implementation to
  use for CUDA Dispatchers.
- Interpose use of the cache in the `CUDADispatcher.compile()` function,
  along with some small refactoring to create the `add_overload()`
  function.

Co-authored-by: stuartarchibald <stuartarchibald@users.noreply.github.com>