Investigate perfomance bottle neck

[ArtDu]

We have a case when the connector slows overall request performance with space.select/insert/replace methods.
We've seen that CPU in Java application achieves the ceil, but Tarantool could achieve more.
In my opinion, we can add a test in jmh that creates mapper for each request like it is in space.select:

1. One instance
2. Empty call with massive response
3. Recreating the mapper for each request
   Ref:

   cartridge-java/src/test/java/io/tarantool/driver/benchmark/BenchmarkRunner.java

   Lines 70 to 75 in 67c9f67

   	TarantoolResult<TarantoolTuple> tuples = plan.tarantoolClient.call(
   	"return_arrays_with_different_types",
   	Collections.emptyList(),
   	plan.defaultMapper,
   	plan.resultMapper
   	).join();