Incomprehensible benmark results from two `dfbench`s built in different dirs

### Is your feature request related to a problem or challenge?

I found a extremely strange different results from two `dfbench` binaries...
- The first one is built by running `cargo build --release --bin dfbench` in `datafusion` dir.
- The second is built by running `cargo build --release --bin dfbench` in `datafusion/benchmarks` dir.

And in my local, I found in `main branch`:
- The first one cost 9200+ms when running q22 in clickbench.
- The second one cost 8800+ms when running q22 in clickbench.

In `my branch` in #11802
- The first one cost 8800+ms when running q22 in clickbench.
- The second one cost 9200+ms when running q22 in clickbench.

### Describe the solution you'd like

I want to know why... It confused me much...

### Describe alternatives you've considered

_No response_

### Additional context

_No response_

Start by reproducing the q22 ClickBench comparison using `cargo build --release --bin dfbench` from both the `datafusion` and `datafusion/benchmarks` directories, including the main branch and the branch from #11802. Trace why the binaries produce different timings and document or correct the cause so the build-location effect is understood.

Written by the indexing model from the issue text.