fix: replace per-token String with byte range

[eikopf]

Summary

• Changed Token struct to store span: Range<usize> instead of text: String, eliminating one heap allocation per token during lexing
• Added Token::text(&self, source: &str) -> &str method to derive token text on demand by slicing the original source string
• Updated Parser to carry the source string and use tok.text(&self.source) instead of &tok.text
• Removed the manually-tracked offset field from Parser; byte offsets are now read directly from token spans

Test plan

• All 566 existing tests pass (cargo test --workspace)
• Lossless round-trip property preserved (token text derived from source matches original)
• Parser error ranges still correct (using tok.span.clone() instead of manual offset tracking)

Closes #16

🤖 Generated with Claude Code