Spike: Improve Generics for Return Types in Primitives

## Problem
- There are certain primitives where the return_type depends on the input_type, such as AddNumeric, MultiplyNumeric, GreaterThan, GreaterThanEqualTo, LessThan
```
input_types = [ColumnSchema(semantic_tags={'numeric'}), ColumnSchema(semantic_tags={'numeric'})]
return_type = ColumnSchema(semantic_tags={'numeric'})
```
- The following situations can happen
- float + float = float
- float + int = float
- int + int = int
- However, since we can't define primitive return types that depend on the input types, this leads to a situation where after calculate_feature_matrix, **we have to re-run inference. This is expensive, should not be necessary.**

## Request
- This spike is to create a design document to possible solutions to this problem.