[C++][Acero] Refine hash join benchmark

[zanmato1984]

Describe the enhancement requested

The current hash join benchmark can be refined in the following aspects:

1. It is using openmp for multi-threading, and this is the solely place that depends on openmp library. Based on my experience of recent benchmarking, openmp can cause tricky inaccurate benchmarking numbers (some obvious improvement shows worse number in one or two out of tens of cases, I still don't know why), and make diagnosis hard (I was using flame graph and openmp makes it hard to interpret, see the following two pictures).
   Flame graph with openmp:
   
   Flame graph without openmp:
   
   I'm not entirely sure why openmp was introduced at the first place (maybe @save-buffer can elaborate a bit?), but I think at this point we can trivially replace it with arrow-native multi-threading primitives. (Besides a bunch of simplification of cmake could be done).

2. In certain benchmarks, we are more interested in the number of rows from the build side. Currently all benchmarks are based on the number of rows from the probe side.

Component(s)

C++

[zanmato1984]

Issue resolved by pull request 45593
#45593