refactor: unify two competing token type systems (issue #215 completion)

[ajitpratap0]

Problem

Tokenizer outputs []models.TokenWithSpan but the parser consumes []token.Token. A 716-line conversion layer (token_conversion.go) exists solely to strip span info — unnecessary architectural duplication.

Impact

• Every parsing operation triggers an O(N) conversion of the entire token stream
• Two parallel token representations create confusion
• The conversion layer has its own keywordBufferPool — a smell indicating the conversion cost is real

Fix

Unify to models.TokenWithSpan directly in the parser. Store positions separately only when ParseWithPositions() is called. Eliminates the conversion step entirely.

Source

Pre-release architecture review (v1.8.0). Continuation of #215.