See this issue: bevyengine/bevy#14302 and the accompanying PR: bevyengine/bevy#14316 (comment) for context and an example implementation.

The motivating problem is this: operating multiple times on a normalized quaternion will accumulate small errors that will eventually result in it no longer being normalized. While the error is still small, a full .normalize() is not necessary; there are cheaper ways to get such a quaternion normalized again.