Optimize ros1

[Carter12s]

Description

• Provides a ~20-30% speedup (high variance) on image benchmark by doing more intelligent pre-allocation

This may have some performance penalties on topics which have frequent small messages and infrequent large messages, but I'm not really sure we'll see a real-world case of that.

Fixes

Closes: number

Checklist

• Update CHANGELOG.md

[Carter12s]

This is a worst-case performance improvement on the image bench, I've done a lot of runs and see high std.dev between runs, but with 10k samples over ~15s I think this is pretty real.

[Carter12s]

I've also generated an equivalent C++ benchmark and run it on the same system the above benchmarking was done on. I've done my best to make the benchmarks as equivalent as possible:

rosrun roslibrust_cpp_image_bench cpp_image_bench --mode driver --iterations 10000 --warmup 50
mode: driver
samples: 10000 measured, 50 warmup
image: 1920x1080 rgb8, payload_bytes: 6220800
mean: 1.228 ms
median: 1.203 ms
p95: 1.377 ms
min: 0.992 ms
max: 4.984 ms
stddev: 0.094 ms
throughput_payload: 4832.606 MiB/s (5.067 GB/s)
json: {"mode":"driver","iterations":10000,"warmup":50,"width":1920,"height":1080,"channels":3,"encoding":"rgb8","queue_size":1,"tcp_no_delay":false,"payload_bytes":6220800,"mean_ns":1227622.784800,"median_ns":1202900.000000,"p95_ns":1376580.000000,"min_ns":991740.000000,"max_ns":4983630.000000,"stddev_ns":93504.184201,"throughput_mib_s":4832.605961,"throughput_gb_s":5.067355}

I have strong confidence that our ROS1 implementation (at least on the case of large image messages) is now as fast, or FASTER than noetic roscpp.