Faster lexical analyzer

[soutaro]

Extracted from #2652

This PR improves the data structure of lexer in RBS.

It improves the parsing performance from ~14 i/s to ~16 i/s measured by bin/benchmark-parse.rb.

I have `gem_rbs_collection` repository too to load `activerecord` rbs files.

Baseline

➜  rbs git:(38724282) bundle exec ruby bin/benchmark-parse.rb core/**/*.rbs ../../ruby/gem_rbs_collection/gems/activerecord/8.0/*.rbs sig/**/*.rbs
Benchmarking parsing 177 files...
✅ 14.506 i/s (±0.000%)

Fix lexer

➜  rbs git:(fix-lexer) bundle exec ruby bin/benchmark-parse.rb core/**/*.rbs ../../ruby/gem_rbs_collection/gems/activerecord/8.0/*.rbs sig/**/*.rbs
Benchmarking parsing 177 files...
✅ 16.667 i/s (±0.000%)

[soutaro]

The lexer data structure now stores the code point of next character, so that peeking next character can be implemented really faster than reading next character from buffer.

[soutaro]

This function assigns the next codepoint in the buffer and it's byte length.

[soutaro]

We assume the character encoding of RBS files is ASCII compatible, like Ruby source file.
If it is ASCII character, it is a single-byte character.

[soutaro]

Another hack to support encoding other than utf-8.
It doesn't know the exact unicode code point of the next character in other encoding, and it returns a random code point instead. Lexer reads the character, but because the random character doesn't have any meaning for lexer, it works perfectly.

We may want to return a upper case character to support multi-byte class/constant names.