Eliminate redundant GPU force reduction event dependency

Eliminate redundant GPU force reduction event dependency

CPU force transfers happen on the same strem as GPU reduction, so no explicit dependency is needed.

Additionally, due to a StatePropagatorDataGpu bug, whenever the force readiness event queried is not AtomLocality::All, an incorrect event is returned leading to circular dependency on force reduction, as described in #4032.

This change however does not fix the StatePropagatorDataGpu bug, but it should help avoiding workarounds for the new SYCL backend (#3932).

Refs #4032 #3988