Since #16173, `account_summaries` and `follow_recommendations` are not
populated at creation time, which causes concurrent refresh to fail.

As we currently only use those materialized views right after explicitly
refreshing them, this commit changes refreshes to not be performed
concurrently. This will fix the issue and ensure the refresh completes
faster while using less resources.

Express follow_recommendations in terms of account_summaries rather than
accounts, integrate filters that are unconditionally used, and materialize
the resulting view.

This should result in the bulk of the computation being performed only once
instead of **once per recommendation language**.