TSDB ingest performance: combine routing and tsdb hashing

[felixbarny]

Instead of hashing dimensions during routing and then again during document parsing, this combines the two steps. The tsid is created during routing and then used to create a routing hash. The tsid is then sent to the data nodes which acts as a signal that creating the tsid during document parsing isn't required anymore.

Instead of populating the index.routing_path setting that can differ from the document dimensions, this now populates a new index.dimensions index setting containing all dimensions. This setting isn't user-configurable (todo). In case users manually set index.routing_path, the new optimization doesn't kick in so that routing and tsid creation is working as before. Additionally, if the dimension fields can't be expressed as a simple set of path matches (for example when using a dynamic template with a match_mapping_type that sets time_series_dimension: true), it falls back to populating index.routing_path.

As an additional benefit, the new _tsids are shorter, which may have benefits at query time. While they're shorter, they still retain the main properties: clustering similar time series together (which helps in compression) and making collisions very unlikely. More details in the JavaDoc of TsidBuilder.

I've added a dependency on hash4j which provides an efficient way to hash strings, without having to create a temporary utf-8 byte array, as well as a nice API.

Remaining issues to work out:

• Make index.dimensions a private setting
• Update index.dimensions when adding a new dimension field to the mappings.
• Ensure that all dynamic templates are incorporated into index.dimensions so that the coordinating node always knows which paths will be considered dimensions.

[elasticsearchmachine]

Pinging @elastic/es-storage-engine (Team:StorageEngine)