Optimise out string allocations and copies in keyword lookup

[alexander-beedie]

Follow-up to #2225.

Description

• Found another ~25 or so places where we can make peek_token → peek_token_ref updates.

• Profiling showed keyword lookups are a significant hotspot, so the following optimisations were made:

  • Added a Token::make_word_owned variant to avoid to_string() copies, factoring out the keyword lookup.
  • The keyword lookup itself was then optimised to eliminate to_ascii_uppercase() heap allocations.

Results

Ran sqlparser_bench several times with a fresh baseline to confirm the speedup 👍
These gains are in addition to the previous optimisation PR (baseline includes them).

┌──────────────────────────────┬───────────┬───────────┬────────┐
│ benchmark                    │ baseline  │ this pr   │ change │
├──────────────────────────────┼───────────┼───────────┼────────┤
│ sqlparser::select            │ 2.691 µs  │ 2.477 µs  │ −8.0%  │
├──────────────────────────────┼───────────┼───────────┼────────┤
│ sqlparser::with_select       │ 12.944 µs │ 11.652 µs │ −10.4% │
├──────────────────────────────┼───────────┼───────────┼────────┤
│ parse_large_statement        │ 4.467 ms  │ 3.880 ms  │ −13.1% │
├──────────────────────────────┼───────────┼───────────┼────────┤
│ format_large_statement       │ 242.62 µs │ 232.23 µs │ −4.1%  │
├──────────────────────────────┼───────────┼───────────┼────────┤
│ select_100_columns           │ 68.56 µs  │ 63.94 µs  │ −6.7%  │
├──────────────────────────────┼───────────┼───────────┼────────┤
│ select_100_qualified_columns │ 142.47 µs │ 133.75 µs │ −6.1%  │
└──────────────────────────────┴───────────┴───────────┴────────┘

(Benchmarks were run on an Apple M3 Max)

[iffyio]

LGTM! Thanks @alexander-beedie!