[BUG] Improve fallback mechanism if there are back to back transitions between the CPU and the GPU

[kuhushukla]

Describe the bug
Consider a query which has 5 operators :

• Scan (NOT_ON_GPU)
• Filter (ON_GPU)
• Project (NOT_ON_GPU)
• Union (ON_GPU)

We will have 2 RowToColumnar and 1 ColumnarToRow transitions.
In cases where there is a wide schema or a significant amount of data, this can cause poor performance for that stage.
This example can be extended to other operators as well, however, we may consider joins etc. might be a bit more subtle on what the performance benefit of the operator will be versus the overhead of R2C.

Steps/Code to reproduce bug
Use above. More to follow.

Expected behavior
We should be smart about not moving things back and forth between the CPU and GPU at least in obvious cases that involve projects and filters. This will limit the cost the jobs incurs from these transitions.

Environment details (please complete the following information)

• Reported for on-prem but applies to any platform

Additional context
Wide Schema with thousands of columns with strings