A performance regression was introduced in
https://gitlab.com/gitlab-org/gitlab/-/merge_requests/38095#note_412661217
. This reverts the relevant code that was necessarily passing through a
limited set of project IDs for Elasticsearch.

This also takes an extra refactoring step. Since we are passing project
IDs through to this method it no longer has the projects which was
necessary for the `#generic_search_results` method. But this was only
ever used for user search because users are not indexed in Elasticsearch
and as such cannot actually search using Elasticsearch. I decided it
kept things simpler to move this logic further up the calling stack so
we don't need the `Elastic::SearchResults` objects to be delegating back
again to the normal `SearchResults` objects. This is much simpler
because the calling code is already responsible for determining whether
or not Elasticsearch should be used anyway.