[P1] Address masking is inconsistent for `NonZeroAddressType` in `arith.cmpi` lowering path

[ngmachado]

Problem

ConvertCmpOp masks both AddressType and NonZeroAddressType, but ConvertArithCmpIOp only masks AddressType.

Evidence

• ConvertCmpOp handles both: src/mlir/ora/lowering/OraToSIR/patterns/Arithmetic.cpp:150
• ConvertArithCmpIOp only checks address: src/mlir/ora/lowering/OraToSIR/patterns/Arithmetic.cpp:740

Impact

Potential high-bit garbage leakage into comparisons through one lowering path.

Acceptance Criteria

• Both compare paths apply identical 160-bit masking rules for all address-like types.
• Add targeted cmp regression for address and non_zero_address through both paths.