perf: Let the lexer operate on u8 instead of char to improve read performance.

[LinZhihao-723]

Description

When reading char, the underlying reader has to do utf8 validation, which is slow and unnecessary in our use case. This PR changes the lexer to directly operate on u8, which increases the performance by ~15% in the benchmark data set.

Validation performed

• Ensure all workflows passed.
• Ensure example works well, with unicode handled
• Ensure benchmark performance gets better