Inconsistent type inference for date add/subtract interval operations

[LeonidChistov]

Discussed in #574

Following operations are defined in Substrait prototypes:

add(date, interval_year): -> timestamp
add(date, interval_day): -> timestamp
subtract(date, interval_year): -> date
subtract(date, interval_day): -> date

It looks like add and subtract behavior need to be consistent with regards to result type.

However, different databases/frameworks have different type inference for these operations:

• Postrgres returns timestamp in all four cases
• Calcite infers return type as date in all four cases

Probably it is not so important which type inference approach will be used by Substrait, because type casts from/to timestamp for operands/result can be used to adjust to specific DB semantics.

But there is no reason to let add and subtract have different type inference.