Decimal multiply kernel may cause precision loss

[viirya]

Describe the bug

Currently decimal multiply kernel (multiply_dyn_decimal) sets original scale by a divide operation. It tries to restore the scale assuming lhs and rhs have the same scale. For example, multiply two decimals (-491, -591) with precision/scale (5, 2), it will divide 290181 by 2 to get the result 29.01 with same precision/scale (5, 2). But by doing this, it loses decimal precision. For example, in Spark, the answer is 29.0181 with precision/scale (11, 4).

To Reproduce

Expected behavior

I'd expect that it behaves like Spark as mentioned above.

Additional context

[mustafasrepo]

I guess you meant -4.91 and -5.91 instead of (-491 and -591).

[viirya]

Yes, -491 and -591 are their i128 representation in the array with precision/scale (5, 2).

[Dandandan]

Also see
#5143