fix: reject overflowing division precision

[yuefdev]

Reject math.MinInt32 before negating division precision. Negating this value as an int32 wraps back to itself, so the existing exponent check can miss the overflow and return an incorrect non-zero quotient.

The regression uses New(1, math.MinInt32) divided by one, without constructing a huge number. Both QuoRem and DivRound fail the regression before the fix. The adjacent valid precision bounds are checked too.

Focused division tests, go vet ./... and formatting pass.

[mwoss]

Thanks, good catch! I pushed a small cleanup on top: the MinInt32 check is now part of the existing overflow check, so there's one panic site. The test moved into TestDecimal_Overflow next to the other overflow panics and uses didPanic. I kept the MinInt32+1 boundary case and dropped the doc line, since the other overflow panics aren't documented either. Same behavior, just less code.