Add Clickbench inspired REGEXP_REPLACE benchmark

[Dandandan]

A new grep-captures benchmark modeling ClickBench Q28 (analytic SQL benchmark), which extracts the URL host from HTTP Referer values:

REGEXP_REPLACE(\"Referer\", '^https?://(?:www\.)?([^/]+)/.*\$', '\1')

Currently rust/regex is about 6.5x slower than javascript/v8 on this benchmark.

results

  Engine                          |   Median |  vs slowest |  vs rust/regex
  --------------------------------+----------+-------------+----------------
  javascript/v8 23.10.0           |  2.48 ms |       21.7x |          2.4x
  python/re 3.13.7                | 12.63 ms |        4.3x |         12.4x
  perl 5.34.1                     | 15.31 ms |        3.5x |         15.0x
  python/regex 2026.5.9           | 16.46 ms |        3.3x |         16.1x
  rust/regex 1.7.3 (no patch)     | 16.43 ms |        3.3x |         16.1x
  rust/regex (patched).           |  1.02 ms |       52.7x |          1.0
  go/regexp 1.21.5                | 18.25 ms |        2.9x |         17.9x
  rust/regex/lite 0.1.8           | 53.73 ms |        1.0x |         52.7x

[BurntSushi]

This is amazing work. The write-up here is especially lovely. Thank you for that.

I see your PR on regex too and I've only had a chance to look at it broadly. So far I don't see any red flags. I love the general idea of looking for load bearing regexes and optimizing them like you've done here. That was my intent with the strategy infrastructure anyway.

With that said... I don't have a good sense of how well the strategy infrastructure scales. This isn't totally fair to pin on you, but could you run the rust/regex compile benchmarks (not Rust compilation, but regex compilation) before and after your change? I just want to check whether the additional strategy checking comes with a regression here. And if so, how big.

Thank you. :-) <3

[Dandandan]

This is amazing work. The write-up here is especially lovely. Thank you for that.

I see your PR on regex too and I've only had a chance to look at it broadly. So far I don't see any red flags. I love the general idea of looking for load bearing regexes and optimizing them like you've done here. That was my intent with the strategy infrastructure anyway.

With that said... I don't have a good sense of how well the strategy infrastructure scales. This isn't totally fair to pin on you, but could you run the rust/regex compile benchmarks (not Rust compilation, but regex compilation) before and after your change? I just want to check whether the additional strategy checking comes with a regression here. And if so, how big.

Thank you. :-) <3

I must admit this was mostly done by Claude (with some bit of steering / cleaning up of course) and I am not familiar with the code base so feel free to let it "cool down" a bit ;) or to let it inspire another implementation

The compilation time is within noise (earlier comment said it was slower (~16 vs ~70 us), but that was compared to rust/regex 1.7.3 which seems to have some lower overhead). 1.12.2 has similar overhead:

  Build                          | extract-host | compile
  -------------------------------+--------------+---------
  upstream master (no patch)     |     14.16 ms | 70.58 us
  my fork (with patch)           |      1.02 ms | 71.62 us
  -------------------------------+--------------+---------