While working on #55713, I realized that the current way of propagating the metadata in the vectorized engine is incompatible with some things we would like to do. As a reminder, we currently buffer up all metadata in MetadataSources and return it when draining the whole flow, but there are cases when we would like to propagate the metadata once it is created:

1. we emit the metadata about the number of rows read from rowexec.TableReaders which is then used in combination with the estimated number of rows to be scanned (which is obtained during planning) in order provide the progress estimate for a query (shown on SHOW QUERIES in phase column)
2. the table statistics collection emit progress metadata as well, and although we don't have the vectorized table statistics collection yet, it'd be nice to support the progress update when we do.

The current thinking is that we probably want to implement a pull-based model that is able to return the buffered meta since the previous call (something like DrainBufferedMeta which - unlike DrainMeta - doesn't have an assumption of being called when the flow is shutdown) and then have the root materializer periodically poll its metadata sources.

Jira issue: CRDB-3634