Benchmark: Concurrent queries in constrainted environment about arrow-datafusion HOT 1 OPEN

Here is a potential strawman implementation that I think would be fairly straightforward to implement:

1. Use criterion and add a new bench benchmark (takes care of the warmup, measurements, etc)
2. Use the tpch / clickbench dataset (created via ./bench.sh data clickbench, etc)
3. 2-3 queries (e.g. a TOP 10 query with predicates, and an aggregate with low cardinality)
4. Use default target-partitions / threadpool (e.g. num cores on machine)
5. The benchmark is how long it takes to complete some fixed number queries (e.g. 100)

Then we could run

# Run q1 100 times, start the next query as soon as the first is complete
concurrency_q1_concurrency1
# Run q1 100 times, running 2 queries at any time
concurrency_q1_concurrency2
# Run q1 100 times, running 4 queries at any time
concurrency_q1_concurrency4

ANd the same for one or two other queries

from arrow-datafusion.