Partially addresses #47424.

We request Gitaly in a N+1 manner to build discussion diffs. Once the diffs are from different revisions, it's hard to make a single request to the service in order to build the whole response.
With this change we solve this problem and simplify a lot fetching this piece of info.

Previously, we used Psych, which would:

1. Check if a string was encoded as binary, and not ASCII-compatible.
2. Add the !binary tag in that case.
3. Convert to base64.

We need to do the same thing, using a new column in place of the tag.

This adds an ID-less table containing one row per file, per merge request
diff. It has a column for each attribute on Gitlab::Git::Diff that is serialised
currently, with the advantage that we can easily query the attributes of this
new table.

It does not migrate existing data, so we have fallback code when the legacy
st_diffs column is present instead. For a merge request diff to be valid, it
should have at most one of:

* Rows in this new table, with the correct merge_request_diff_id.
* A non-NULL st_diffs column.

It may have neither, if the diff is empty.