Lexical analysis for SQL text.

Mirrors SQLite's tokenize.c, which classifies bytes through a 256-entry lookup table and resolves keywords with a generated perfect hash. In Elixir the same job is done with binary pattern matching: each clause of scan/3 corresponds to a character class, and keyword recognition is a MapSet lookup on the downcased identifier.

Tokens are {type, value, line} tuples:

• {:keyword, :select, 1} — case-insensitive SQL keywords
• {:id, "users", 1} — identifiers (bare, "quoted", [bracketed], or backticked)
• {:int, 42, 1} / {:float, 3.14, 1} — numeric literals
• {:string, "abc", 1} — single-quoted string literals ('' escapes a quote)
• {:blob, <<0xAB>>, 1} — x'AB' hex blob literals
• punctuation and operators such as {:comma, ",", 1}, {:le, "<=", 1}