These tests will fail if elastic/ml-cpp#2376 with
them unmuted. elastic#88901 will follow up with the Java
side changes.

These tests will fail if elastic/ml-cpp#2376 with
them unmuted. #88901 will follow up with the Java
side changes.

…8901)

This change will facilitate a performance improvement on the C++
side. The request ID and cache hit indicator are the parts that
need to be changed when the C++ process responds to an inference
request. Having them at the top level means we do not need to
parse and manipulate the original response - we can simply cache
the inner object of the response and add the outer fields around
it when serializing it.

Companion to elastic/ml-cpp#2376

In elastic#2376 the results writer for inference was changed
to splice a preformatted cached section of the final result
document into an outer wrapper that varies per request.

The approach used worked when assertions were disabled but
tripped assertions in RapidJSON due to its internal counters
thinking that the spliced portion was a single value instead
of a key and value together.

This PR fixes the code when assertions are enabled by
taking advantage of the fact that the internal value
count of the RapidJSON writer is accessible to derived
classes, so we can add a method that adds a key and
value together and increments the counter twice.

Adapt the test script for the result format changes in #2376