fix: stream remote io paths and benchmark decompression backends

[digizeph]

Summary

• make async downloads preserve raw bytes and stream cache population instead of buffering entire payloads in memory
• reuse default blocking HTTP clients across GET and HEAD paths
• replace string-parsed S3 status handling with a structured status error and stream S3 reads through a bounded channel
• add gzip backend and bz2 decompression benchmarks plus a helper script to run the comparison matrix
• update changelog, docs, and tests for the new behavior

Benchmarks

16 MiB generated text corpus, measured on both the raw decoder path and oneio::get_reader.

• cloudflare-zlib: raw 1.162 GiB/s, oneio 1.135 GiB/s
• zlib-rs: raw 1.035 GiB/s, oneio 1.017 GiB/s
• zlib-ng: raw 1.020 GiB/s, oneio 1.014 GiB/s
• miniz_oxide: raw 686 MiB/s, oneio 680 MiB/s
• bz2: raw 76.8 MiB/s, oneio 76.9 MiB/s

cloudflare-zlib was the fastest gzip backend in this run. zlib-rs and zlib-ng were close behind and materially faster than miniz_oxide.

Example

Run the backend comparison locally:

./scripts/bench_decompression_backends.sh

Behavior note:

oneio::download_async("https://example.com/data.csv.gz", "local_data.csv.gz").await?;

The async download now preserves the remote bytes, matching download().