Use Total Ordering for Aggregates and Refactor for Better Auto-Vector…

…ization (#5100)

* Refactor numeric aggregation kernels to make better use of
auto-vectorization.

Remove the explicit simd implementations since the autovectorized
versions are faster on average.

The min/max kernels for floating point numbers now use the total order
relation.

* Comments and cleanup

* Clippy fixes

* Use largest/smallest bit patterns for float MIN/MAX constants, these differ from the canonical NAN bit pattern

* Add test coverage for aggregating large non-null and float inputs

* Add test with negative NaN

* Rename MIN/MAX constants to make it explicit they use the total order relation