fix(primitives): return Err on parse_units multiply overflow

[Kropiunig]

Motivation

ParseUnits::parse_units returns a Result and is documented/tested to reject out-of-range inputs with an Err — see the existing assert!(parse_units("1", 80).is_err(), "overflow") test. It correctly guards the 10^exponent factor with checked_pow, but then folds that factor into the parsed mantissa with an unchecked *=:

let mut a_uint = U256::from_str_radix(amount, 10)?;
a_uint *= U256::from(10)
    .checked_pow(U256::from(exponent - dec_len))
    .ok_or(UnitsError::ParseSigned(ParseSignedError::IntegerOverflow))?;

When 10^exponent fits in the target integer (so checked_pow succeeds) but mantissa * 10^exponent does not, the overflow is silently swallowed:

• Unsigned U256: ruint implements Mul/MulAssign via wrapping_mul, so the multiply wraps to a wrong value in every build profile. For example parse_units("2", 77) — 10^77 < U256::MAX < 2·10^77 — returns Ok(<wrapped garbage>) instead of Err.
• Signed I256: Signed::mul uses debug_assert!(!overflow), so the equivalent input panics in debug builds and silently wraps in release. For example parse_units("-6", 76) (10^76 < I256::MAX < 6·10^76).

Both outcomes contradict the intended contract (graceful Err on overflow).

Solution

Replace the unchecked *= with checked_mul on both the unsigned and signed paths, propagating the same UnitsError::ParseSigned(ParseSignedError::IntegerOverflow) already used for the checked_pow guard. Minimal change, no new error variants.

Adds a regression test (test_parse_units_mul_overflow) covering the unsigned wrap and signed overflow paths. It fails before this change (unsigned returns Ok; signed panics in debug) and passes after.

Testing

cargo test -p alloy-primitives --lib utils::units — all pass. cargo +nightly fmt --check clean.