Use modular exponentiation for concrete EXP

[DicksonWu654]

Summary

• Replace the concrete lhs ** rhs shortcut with three-argument pow modulo the bit-vector width.
• Preserve all symbolic exponentiation paths.
• Add ordinary wrapping/equivalence cases and a maximum-width exponent regression.

Rationale

EVM EXP results are reduced modulo 2^256. Constructing the entire Python power before bit-vector truncation is equivalent but can allocate an enormous intermediate. Modular exponentiation computes the same result while keeping intermediates bounded; using 1 << size also preserves the helper's generic bit width.

Validation

• Static diff/whitespace validation passed.
• Added focused direct tests for ordinary, wrapping, and very large exponent cases.
• Executable tests, builds, package-manager commands, and hooks were not run locally; repository CI is expected to provide executable validation.