`normalized()` panics on debug and returns a wrong value on release for large negative scales

[DRMacIver]

use bigdecimal::BigDecimal;
use num_bigint::BigInt;

fn main() {
    let d = BigDecimal::new(BigInt::from(10), i64::MIN);
    println!("input = {}", d);
    let n = d.normalized();
    println!("normalized = {}", n);
}

On a debug build this panics:

input = 10e+9223372036854775808

thread 'main' panicked at src/lib.rs:926:21:
attempt to subtract with overflow

On a release build there is no panic, but the result is wrong:

input = 10e+9223372036854775808
normalized = 1E-9223372036854775807

10E+9223372036854775808 and 1E-9223372036854775807 are not the same value; normalized() is supposed to only strip trailing zeros while preserving the value.

Tested on bigdecimal 0.4.10.

BTW, this bug was found using hegel. Happy to contribute the tests if you're interested.