Describe the bug
Pointed out by @jorgecarleitao and @b41sh on #779 (comment)

Semantically, the numerics of an i128 are not the same as the numerics of (months,days,nanos) since i128 + i128 != (months, days,nanos) + (months, days,nanos).

The consequence of defining this type numerically here is that arithmetic kernels will accept this type, but they will yield a semantically incorrect result (e.g. i128 + i128 to sum two intervals of 1 month each).

To Reproduce
Do a numerical operation on two arrays of MonthDayNano type (for example add or subtract them). The arithmetic will be performed directly on the i128 representation rather than field by field

Expected behavior
The intervals should be calculated field by field; So for example to add two MonthDayNano fields, the kernels should add the nanoseconds, days and months separately

Additional context