Improve appender benchmarks

- ensure setup code does not affect benchmarks numbers (this seems to be the case now according the cpu profile)
- do not send the same body as it would make compression more efficient that it actually is
- add separate benchmark for retry behaviour

The issue does not name benchmark files or tests. Locate the appender benchmark entry points and inspect the CPU profile to separate setup from measured work, then check how request bodies are generated. Done means setup is excluded, bodies avoid compression bias, and retry behavior has its own benchmark.

Written by the indexing model from the issue text.