Understanding and reducing compilation overheads

The compiler seems to have non negligible overhead when dealing with large graphs (for very large widths, e.g., 200). It would be good to understand these overheads and determine if they can be reduced.

I would try to compile the nfa-regex script (https://github.com/binpash/pash/blob/main/evaluation/benchmarks/oneliners/nfa-regex.sh) with increasing widths to estimate overheads. The `--dry_run_compiler` flag could be used to not execute the compiled script but still compile it (it is pretty old so I am not sure if it still works).

The first step would be to understand which method calls of the compiler contribute mostly to the compilation overhead. The next step would be to reimplement some of them to minimize such overheads.

Start with evaluation/benchmarks/oneliners/nfa-regex.sh and compile it at increasing widths using the --dry_run_compiler flag. Identify which compiler method calls dominate compilation overhead, then measure whether reimplementing selected calls reduces that overhead.

Written by the indexing model from the issue text.