[FEA] Support partial fallback to the CPU on Databricks for percentile

[ttnghia]

Somehow the plugin cannot replace Percentile class on Databricks when the replace mode is set to final|complete. As such, we cannot support partial fallback to the CPU.

The corresponding test (in #9296) needs to be xfail while the root cause needs to be investigated and tracked by this issue.

[revans2]

I did some quick tests, and the plugin is complaining that

 cannot run on GPU because mixing CPU and GPU aggregations is not supported

for the first SortAggregateExec, but all of the children show that they are on the GPU. I'll try to dig a little deeper.

[revans2]

Okay I traced it down. We put in some code for databricks to prevent partial replacement in some cases and it is hitting here

#4963

is the issue to fix this. So we can decide if we are hitting this in error or not.