RUBY-3946 - Check connection back in when a load-balanced initial command fails

[drymar]

Fixes https://jira.mongodb.org/browse/RUBY-3946

In load-balanced topology, cursor-returning operations check a connection out of the pool before executing the initial command so the cursor can retain it. When that command fails, no cursor exists to drain and check the connection back in, so the pool permanently loses a slot on every failure; a process accumulating such failures eventually cannot check out any connection (ConnectionCheckOutTimeout with the full pool checked out and nothing pending) and never recovers.

ChangeStream and Database#cursor_command already release the connection on failure (the former since #3041). This applies the same handling to the remaining initial-command sites:

• find (Iterable)
• aggregate (Aggregation)
• mapReduce (MapReduce)
• parallel scan getMore (Readable)
• listCollections (Database::View)
• listIndexes (Index::View)

Connections pinned to a transaction are left checked out, matching Database#cursor_command.

Reproduction:

pool = client.cluster.servers.first.pool
client[:foobar].find(_id: BSON::ObjectId.new).to_a # warm up
pool.available_count #=> 5 (baseline)
3.times { client[:foobar].aggregate([{ '$foo' => 1 }]).to_a rescue nil }
pool.available_count #=> 2
GC.start; sleep 12  # finalizers + reaper
pool.available_count #=> 2, slots never recovered

[comandeo-mongo]

@drymar Thank you for your contribution, great work!