Implement dot-product in a compute kernel

*Note*: migrated from original JIRA: https://issues.apache.org/jira/browse/ARROW-11115

An efficient implementation of the [dot-product|https://en.wikipedia.org/wiki/Dot_product] is useful for many machine-learning applications.

I would propose to treat null as zero for the implementation. This behavior is sensible because it corresponds to dropping any observation where null-data is present.