perf(rust): update string replacement codepaths following new benchmarking

[alexander-beedie]

Prepare for graphs :)

TLDR

• Currently we leave the small-string regime for str::replacen, and use the regex engine for all other literal replacements (where len(s) > 32).
• Benchmarking the latest code shows that the regex-based replace (in conjunction with NoExpand) now outperforms in essentially all string regimes (from small to large), allowing us to both simplify the code and get a bit of extra performance in the small-string regime.

Build

Compiled with:

maturin develop --release -- -C codegen-units=16 -C lto=thin -C target-cpu=native

Compared

(regex)       =>  reg.replace(s, NoExpand(val))
(str.replace) =>  Cow::Owned(s.replacen(pat, val, 1))

Results

Tested on strings between 8 - 4096 chars in length, replacing between 0% and 100% of the target string, averaging a large number of repeat runs to reduce timing noise...

polars_str_replace_bench.xlsx

[alexander-beedie]

Can replicate the benchmark locally using the following script:
https://gist.github.com/alexander-beedie/a732e4213709696c8d603294d2e1c25c

[alexander-beedie]

In light of the new results, I'll revisit contains next, as it uses a similar optimisation.
Update: done #6811.

[ritchie46]

Nice! We can thank this PR I believe: rust-lang/regex#930