GeezerVison

…IMAL ROUND

- DynamicDecimal multiplication and the 256-bit branches of addition added
  the digits dropped from the result to the exponent instead of subtracting
  them.  CAST(DECIMAL AS INTERVAL DAYS) and MAKE_TIME with many fractional
  digits gave small wrong values.
- The conversion of a DynamicDecimal with more than 38 fractional digits to
  DOUBLE panicked, e.g. CAST(PARSE_JSON('1e-100') AS DOUBLE).
- LongInterval::years negated the months, which overflows for the minimum
  interval.
- ROUND and BROUND of a DECIMAL panicked when rounding more than 38 digits
  away from the last digit of the scale.

Signed-off-by: Mihai Budiu <[email protected]>

- ABS of the minimum integer or interval, and unary minus of the minimum
  integer now report an overflow.

Signed-off-by: Mihai Budiu <[email protected]>

…gits

10^digits is infinity or 0 for a large digits argument: ROUND(1.5e0, 400)
was NaN, and ROUND(1e300, 20) was infinity.

Signed-off-by: Mihai Budiu <[email protected]>

Signed-off-by: Mihai Budiu <[email protected]>

Adding two DynamicDecimal values whose exponents differed by more than
38 panicked, because the value with fewer fractional digits was scaled
by a power of 10 that does not fit in an i128.  The datagen connector
adds such values for a DECIMAL range bound with more than 38 fractional
digits.  The sum now keeps the digits of the other value down to 37
places past the last digit of the first, and is truncated toward zero.

Signed-off-by: Mihai Budiu <[email protected]>