Bugfix/vectorized sum fact kernels
Fix some bugs for vectorized sumfact kernels:
- Add predicates to inoutkeys
- Fix bug in realize_direct_input
Realize direct input
Consider the following example: Three scalar SF kernel with input keys ['a', 'a', 'b']. In this case we have three input interfaces. The algorithm used to divide by two and round down to get the interfaces for the lower and upper part. In this keys this would lead to
lower: ['a',] upper: ['a', 'b']
Instead we need to round up (as long as we only do padding at the end.