Optimisations - Round 3

[lquerel]

This PR is a massive update of the underlying infrastructure used to encode/decode OTLP to/from OTLP Arrow messages.
In the previous version the AdaptiveSchema was used to dynamically adapt the schema for fields that can be either a dictionary or a primitive/binary type (e.g. Int32, String, Binary).

This PR introduce a more general system called RecordBuilderExt (ext for extension to the standard Arrow RecordBuilder). The RecordBuilderExt interprets special annotations attached to the Arrow Schema of OTEL entities. There are currently 2 types of annotation supported:

• optional: used to qualify a field as optional
• dictionary: used to qualify a field as a dynamic dictionary able to detect index overflow and react accordingly either by selecting a better type for the index or by fall-backing to the value type of the dictionary.

This modification aims to improve the performance (CPU, memory allocation, compression) of small batches (<100) and to some extent improve a bit the performance of bigger batches.

Note: On the decoder side, we treat all the fields as optional as in protobuf. Logically some fields are mandatory but in fact, at the protobuf level, nothing prevents them from being optional. A default value will be assigned to any missing field (as in pdata).

A new set of unit tests have been added to track dictionary overflow for metrics, logs, and traces.

Updated benchmark results have been included into the README.md file.

[jmacd]

I understand now how schema building and transformation works at the high level. Looks good. I noticed a few places where panic makes sense now. I filed #121 to make sure the collector won't die in case of panics.