PHPLIB-651: Support $merge and $out executing on secondaries

[jmikola]

https://jira.mongodb.org/browse/PHPLIB-651

POC for mongodb/specifications#1062

[alcaeus]

LGTM pending approval of mongodb/specifications#1062

[jmikola]

When I recently spoke to @kevinAlbs about this (re: mongodb/specifications#1062 (comment)), I mentioned an alternative implementation that could bypass a second server selection attempt (i.e. mongoc_client_select_server) by instead iterating on Manager::getServers() to select a primary.

Thanks to the first server selection attempt, we can be guaranteed that the topology has been discovered. getServers() shouldn't return an empty array unless there really are no servers available. But I also don't think there's much benefit to that approach vs. simply invoking server selection a second time. The edge case is that if there really isn't a primary available, we'll block for serverSelectionTimeoutMS on this second attempt.

But if the root concern is potentially blocking the application for 2 * serverSelectionTimeoutMS, let's consider how that might actually happen:

• If the first attempt is looking for a secondary to execute an $out/$merge pipeline and the topology doesn't have a secondary at all, the default scenario will abort early thanks to serverSelectionTryOnce and abort the entire aggregate() method.
• If serverSelectionTryOnce=false and a loop is permitted, server selection could theoretically discover a secondary just before serverSelectionTimeoutMS expires (e.g. at 29 seconds in) and return that server.
• If we then discover the selected server is a pre-5.0 secondary, we'll re-enter server selection to look for a primary. Worst case, this could also take serverSelectionTimeoutMS to ultimately select a primary (e.g. wait through an election) or hit a server selection timeout.

That scenario seems quite rare, so I'm inclined to keep the code as-is rather and punt on any optimization until we actually have evidence that this causes problems down the line. IMO, the aggregate() helper is already very convoluted so anything we can do to avoid introducing more complexity is welcomed.