New CUDA context handling

For !84 (merged), and revise of !66 (merged)

• Introduce ContextRef struct to handle non-owned reference to CUcontext
  • CUDA API allows us to get CUcontext from pointers to CUDA managed memory, which is needed by !84 (merged)
• Rename Context into ContextOwned, and Arc<Context> into Context
• Take a reference to Arc instead of moved value:

  fn create(ctx: Arc<Context>) { ... }

  new API based on Context = Arc<ContextOwned>

  fn create(ctx: &Context) { ... }

• Introduce accel_derive::Contexted proc-macro to derive updated Contexted trait