Use separate tokio runtime for IO in datafusion-cli

[alamb]

Which issue does this PR close?

• Related to Introduce morsel-driven Parquet scan #20529

Rationale for this change

While profiling #20820 I noticed that datafusion-cli is still not keeping its CPU entirely busy for unclear reasons

One interesting thing I noted is that there is no separate IO threads, even though there should be.

I want to test and see if using a separate IO runtime makes a difference

What changes are included in this PR?

Add a separate IO runtime to datafusion-cli and ensure that all IO happens on that stream

Are these changes tested?

Are there any user-facing changes?

[alamb]

This didn't seem to make a meaningful difference in performance -- I need to look more carefully at the io pattern

[Dandandan]

no separate IO threads

My current view: I think we currently are using spawn_blocking fine to keep it of the main / query threads, which is not optimal (because we create too much/unbounded threads), but OK from separating IO calls / CPU.

I think though that we:

• Don't perfectly balance work - that's where morselization will help to balance work evenly across threads (i.e. all threads are 90% busy most of the time
• Don't interleave IO and CPU - that's where prefetching might help to get the last ~10-15% or so extra utilization
• There is also some IO-level efficiency improvements, like file open / close on each read, which might help a bit (although it will matter even less when it is prefetched).