Prepare async index on merge_request_metrics for deploy metrics

What does this MR do and why?

Deployments::UpdateEnvironmentWorker runs an UPDATE on merge_request_metrics on every successful production deployment. It filters not-yet-deployed merge requests by target_project_id, target_branch, and a merged_at range, but no index supports the selective part, so the planner scans every merged MR on the branch. In a recent GitLab.com slow-query sample it was the single biggest contributor (about 6,066 seconds of database time).

This MR prepares a partial async index:

CREATE INDEX index_mr_metrics_on_project_merged_at_not_deployed
  ON merge_request_metrics (target_project_id, merged_at)
  WHERE first_deployed_to_production_at IS NULL;

merge_request_metrics is over_limit, so the partial WHERE first_deployed_to_production_at IS NULL predicate matters: it covers only not-yet-deployed rows, and a row leaves that set permanently once deployed, keeping the index small.

This is the async-prepare step. A follow-up MR will create the index synchronously with add_concurrent_index and drop the near-useless index_merge_request_metrics_on_first_deployed_to_production_at to keep the table's index count flat.

Issue: #628635 (closed)

Migration output

Up:

main: == 20260910120000 PrepareAsyncIndexMrMetricsOnProjectMergedAtNotDeployed: migrating
main: == 20260910120000 PrepareAsyncIndexMrMetricsOnProjectMergedAtNotDeployed: migrated (0.1173s)

Down:

main: == 20260910120000 PrepareAsyncIndexMrMetricsOnProjectMergedAtNotDeployed: reverting
main: == 20260910120000 PrepareAsyncIndexMrMetricsOnProjectMergedAtNotDeployed: reverted

Prepared index definition recorded in postgres_async_indexes:

CREATE INDEX CONCURRENTLY "index_mr_metrics_on_project_merged_at_not_deployed" ON "merge_request_metrics" ("target_project_id", "merged_at") WHERE first_deployed_to_production_at IS NULL