Optimize lexer, parser, and interpreter for speed

[marksmayo]

Summary

Reduces per-expression overhead across the compile and evaluation pipeline with no change in behavior. Benchmarks in tests/compliance/benchmarks.json improve 8–97% across lex/parse/search with zero regressions.

Changes

• lexer: scan with a local index and slice identifier/number/string runs instead of char-by-char accumulation; consume whitespace runs; find-based scanning for quoted strings and literals.
• parser: precompute per-class nud/led dispatch tables, inline token access and AST node construction on the hot path, and look up the compile cache without raising on a miss.
• interpreter: share a pre-warmed default TreeInterpreter for default-options searches, cache resolved visit_* handlers, use a fixed two-argument visit signature, and short-circuit field access on None.
• functions: resolve signature types to Python types once at registration and reuse the cached result during validation.

Compatibility

Public API (search, compile, Options) and subclass extension points are preserved: overriding TreeInterpreter.visit/_is_false, Functions._validate_arguments, base Visitor instance-level/staticmethod handlers, and Parser._error_nud_token all continue to work. The shared default interpreter's method cache is pre-warmed so concurrent default searches only read it (safe on free-threaded CPython builds).

Testing

• Full test suite: 991 passed, 1 skipped
• Hypothesis property tests: 7 passed
• A/B benchmarks vs. base: 0 regressions, all metrics faster