Introduce a BulkIndexerPool abstraction

[marclop]

Currently, the Appender creates bulkIndexers at runtime when the Appender is created. Instead of doing that, we could pass in a BulkIndexerPool, which contains a bounded pool of bulkIndexers and could be shared by multiple Appender instances to create a maximum memory boundary.

For fairness across different go-docappender instances, it would be good also to implement a per-go-docappender limit so one appended isn't potentially consuming all the bulk indexers.