Optimize benchmarks

[preinlein]

What does this PR do?

Improves benchmark performance and accuracy by implementing iter_batched for all payload generation benchmarks. This approach properly separates setup costs from the actual benchmark measurements, ensuring more accurate throughput metrics.

Motivation

The previous benchmark implementation was mixing setup costs (RNG initialization, payload generator creation, buffer allocation) with the actual serialization costs being measured. Using BatchSize::PerIteration with iter_batched ensures we're only measuring the actual serialization performance.

Additional Notes

• Reduced setup benchmark measurement time from 10s to 5s to speed up the benchmark suite
• Removed 1000 MiB benchmarks for some of the more complex payload types (OpenTelemetry, Fluent, Trace Agent) to keep benchmark runtime reasonable
• Renamed benchmark groups from *_all to *_throughput to better reflect what's being measured
• Fixed a trailing semicolon inconsistency in the setup benchmarks

[preinlein]

• Example optimization hunt #1762
• Update hunt to use targetted benchmarks #1761
• Optimize benchmarks #1760 👈 (View in Graphite)
• Split benchmarks into individual targets #1759
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[preinlein]

The stack can be validated by looking at the final Claude hunt results here: https://drive.google.com/file/d/1mbgFAFuC5s2HwM3IhoNCgqZcmt7Ho7a2/view?usp=sharing