perf(s3): use spawn-based fetcher for S3 block source

[sprites0]

Summary

• Override collect_blocks on S3BlockSource with tokio::spawn batches instead of the default buffered() stream
• Avoids single-task polling bottleneck, yielding ~10x speedup on fast networks (4s -> 3.1s in slow network, -> 0.3s in fast network per 10K blocks in testnet)
• Keeps buffered() as the default trait impl (better for filesystem and other sources)
• Both approaches result in the same number of file descriptors — hyper's connection pool opens sockets based on concurrent requests regardless of whether they're driven by buffered() or tokio::spawn

Test plan

• Verify S3 block fetching performance on fast network
• Verify local/hl_node block sources are unaffected (still use default buffered())

🤖 Generated with Claude Code