Fix signed/unsigned comparison and arithmetic semantics in C codegen

[Copilot]

C's implicit unsigned-promotion rule causes wrong results when mixing signed/unsigned integers — e.g. (int32_t)-1 < (uint32_t)1 evaluates to false because -1 becomes UINT_MAX. Loopy's generated code was subject to this, diverging from Python semantics.

Changes

loopy/target/c/codegen/expression.py — ExpressionToCExpressionMapper

• map_comparison: When operands are integers of opposite signedness, cast both to the narrowest common signed type (via np.result_type; falls back to int64 for int64/uint64).
• map_product (new override): When the inferred result is a signed integer and any factor is unsigned, cast those factors to the result type. Fixes subtraction's a + (-1)*b form.
• map_sum (new override): Same treatment for addends — prevents an unsigned term from silently converting signed co-addends to unsigned arithmetic.
• _is_mixed_sign_integer_pair / _common_signed_dtype (new helpers): Detect mixed-signedness and derive the appropriate signed cast target.

Before / after for int32 sv = -1, uint32 uv = 1:

// Before — wrong
result = ((sv < uv) ? 1 : 0);          // sv promotes to uint32 → UINT_MAX > uv
diff   = sv + -1 * uv;                 // unsigned wrap-around

// After — correct
result = (((int64_t)(sv) < (int64_t)(uv)) ? 1 : 0);
diff   = sv + -1 * (int64_t)(uv);

test/test_c_execution.py

• test_mixed_sign_comparison: All six comparison operators (<, <=, >, >=, ==, !=) across int8/uint8, int16/uint16, int32/uint32, int64/uint64 pairs, three representative value combinations each.
• test_mixed_sign_subtraction: Both subtraction directions (sv - uv, uv - sv) for the same dtype pairs.