[FEA] `cudf::scalar` cache `is_valid` to reduce amount of stream syncs

**Is your feature request related to a problem? Please describe.**
As discovered in https://github.com/rapidsai/cudf/pull/8004 the `is_valid` call in `cudf::scalar` is not cached locally, and therefore a stream sync / `cudaMemcpyAsync` will be executed for each call.

If algorithms such as `clamp` this cost is very high and has a impact on total runtime.

**Describe the solution you'd like**
Cache the value of `is_valid` on initial construction and any modification.