cuda.cooperative: Add BlockScan::ExclusiveScan overloads for single input element per thread

[shwina]

Currently, cuda.cooperative supports ExclusiveScan, but only provides the overloads for multiple data elements per thread.

We should also provide overloads for block-wide exclusive scan where each thread contributes a single element.

In particular, it would be good to start with the following:

• BlockScan::ExclusiveScan 'void (T, T &, ScanOp)
• BlockScan::ExclusiveScan 'void (T, T &, ScanOp, BlockPrefixCallbackOp &)

Later on, we could also support:

• BlockScan::ExclusiveScan 'void (T, T &, ScanOp, T &)
• BlockScan::ExclusiveScan 'void (T, T &, T, ScanOp)
• BlockScan::ExclusiveScan 'void (T, T &, T, ScanOp, T &)

It might be useful to refer to #2691 as an example for how to add overloads to an existing algorithm.