perf: optimize substring (up to 10x faster)

[andygrove]

Which issue does this PR close?

N/A

Rationale for this change

Optimize existing expression.

What changes are included in this PR?

Replaced Arrow's substring_by_char with an implementation that skips the per-element chars().count() (unused for non-negative starts) and adds a bulk ASCII fast path that slices windows by byte offset without decoding UTF-8.

How are these changes tested?

Existing tests + new unit tests.

Benchmark (criterion):

• substring_ long ascii tail: 90.315% faster (base 1049866ns -> cand 101681ns)
• substring_ long ascii prefix: 45.268% faster (base 161701ns -> cand 88502ns)
• substring_ short ascii: 69.169% faster (base 89039ns -> cand 27451ns)
• substring_ long utf8 tail: 34.768% faster (base 3226354ns -> cand 2104613ns)
• substring_ long utf8 prefix: 63.96% faster (base 365291ns -> cand 131650ns)

Full criterion output:

substring: short ascii  time:   [27.371 µs 27.413 µs 27.457 µs]
                        change: [−69.325% −69.169% −69.016%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 10 outliers among 100 measurements (10.00%)
  6 (6.00%) low mild
  2 (2.00%) high mild
  2 (2.00%) high severe

substring: long ascii prefix
                        time:   [88.107 µs 88.511 µs 88.914 µs]
                        change: [−45.520% −45.268% −45.006%] (p = 0.00 < 0.05)
                        Performance has improved.

substring: long ascii tail
                        time:   [101.42 µs 101.97 µs 102.50 µs]
                        change: [−90.371% −90.315% −90.260%] (p = 0.00 < 0.05)
                        Performance has improved.

substring: long utf8 prefix
                        time:   [131.49 µs 131.61 µs 131.77 µs]
                        change: [−64.048% −63.960% −63.868%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 9 outliers among 100 measurements (9.00%)
  3 (3.00%) high mild
  6 (6.00%) high severe

substring: long utf8 tail
                        time:   [2.1015 ms 2.1046 ms 2.1087 ms]
                        change: [−34.887% −34.768% −34.620%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 13 outliers among 100 measurements (13.00%)
  1 (1.00%) low mild
  3 (3.00%) high mild
  9 (9.00%) high severe

[andygrove]

I created an upstream PR in arrow-rs as well: apache/arrow-rs#10334

[mbutrovich]

Thanks @andygrove! Thanks for also upstreaming!

[mbutrovich]

This feels more complex than what I'm used to seeing out of an #[inline]. However, it only has one call site so I'm not terrified of code size explosion.