[BUG] Investigate multiple calls to cudf::rolling_window() from GpuWindowExec

[mythrocks]

This was noticed tangentially, when investigating cudf/pull/7568.

The minimal test-case that reproduced the libcudf issue above consisted of a 0.5KB Parquet dataset containing 15 records across 5 groups.

grouped_rolling_window() seemed invoked once per group, instead of once for the entire column, likely in a bid to keep groups whole. If groups could be packed into larger inputs for grouped_rolling_window(), the performance should be far better.

[jlowe]

@mythrocks Is this still relevant?

[revans2]

@mythrocks Is this still relevant?

We could save some time in building the full offsets that are passed to the underlying window operations. But we have not even measure how much of the time that is taking up.

[mythrocks]

Sorry I missed this.
I didn't actually see a slowdown, just that there seemed to be multiple calls coming through.

I can close this issue for now, and reopen if we find a slow path here.

[revans2]

I want to keep it open because we know that there is duplicate code being called. It may be small, but I want to preserve this because at some point we are going to want to go through the backlog and start fixing things there. Just unassigned yourself from this for now.