Amend BackfillMergeRequestDiffCommitsToPartitioned

Fix batched background migration BackfillMergeRequestDiffCommitsToPartitioned.

Two problems are addressed:

1. NULL column violations failing entire sub-batches

Rows with NULL commit_author_id, committer_id, or sha in the source table caused NOT NULL violations when inserting into merge_request_commits_metadata (see comment). Because each sub-batch runs as a single SQL statement, one bad row failed the entire sub-batch.

2. Full table scans on GitLab.com due to view batching

On GitLab.com the migration batched over SQL views, but PostgreSQL cannot collapse row-style (multi-column) index conditions through a view, causing full table scans that made batching very slow (see comment).

Changes

• The filtered_diff_commits CTE now excludes rows with NULL commit_author_id, committer_id, or sha before either INSERT runs.
• Both the batching strategy and sub_batch_relation now query merge_request_diff_commits directly instead of the views; the cursor bounds stored on each BBM record continue to isolate the four parallel workers by range.
• A fix migration (FixBackfillMergeRequestDiffCommitsToPartitioned) updates existing BBMs in place: sub_batch_size 1000 → 10000 and batch_size 50000 → 500000 to take advantage of the faster queries.
• total_tuple_count is also backfilled for GitLab.com since views have no pg_class cardinality (based on this pattern).
• The original queue migration is left unchanged since this fix ships in the same milestone, and the migration has not yet started on self-managed instances. On GitLab.com, failed jobs will be retried with the new code before the migration finalizes.

References

Related to #527230+

Original MR: Backfill partitioned `merge_request_diff_commit... (!224072 - merged)

MR acceptance checklist

Evaluate this MR against the MR acceptance checklist. It helps you analyze changes to reduce risks in quality, performance, reliability, security, and maintainability.