sql/schemachanger: slow rollout of distributed merge in Cloud via RAFA

[rafiss]

Summary:
Use RAFA to do a slow rollout of the bulkio.index_backfill.distributed_merge.mode cluster setting on Cloud clusters during the 26.3 dev cycle. This will give us real-world signal on whether the feature works correctly before switching it on by default.

From the 26.2 pre-mortem.

Cluster setting: bulkio.index_backfill.distributed_merge.mode

• Current default: disabled
• Target value for rollout: enabled
• Defined in pkg/sql/backfill/distributed_merge_mode.go

Precedent:
We followed a similar process for sql.defaults.autocommit_before_ddl.enabled in #133180 and https://github.com/cockroachlabs/rafa-production/pull/512.

Rollout plan:

1. Wait for a 26.2 patch release to be deployed in CockroachDB Cloud.
2. Ensure the SLI dashboard charts are in place (see related issue for dashboard work).
3. Create a RAFA PR to set bulkio.index_backfill.distributed_merge.mode = 'enabled' on an initial set of high-risk clusters.
4. Monitor the SLI dashboard and Datadog logs for the setting change.
5. If no issues after ~1 week, expand to a broader set of clusters.
6. Once confident, change the hard-coded default from disabled to enabled in pkg/sql/backfill/distributed_merge_mode.go.

Next Steps:

• Create the RAFA PR targeting an initial cluster set
• Monitor for regressions after each rollout stage
• Change the hard-coded default once rollout is validated

Jira issue: CRDB-62441

Epic CRDB-62564

[spilchen]

I think this needs to wait until 26.3 is out. There are some fixes that should be in before we start the slow rollout (namely #167918). This also gives us a good release cycle to find any bugs that are detected due to the metamorphic setting (#167786). Plus 26.3 is the innovation release.