ngx_mzstd moves zstd compression to NGINX thread pools

The Rust-based ngx_mzstd module compresses zstd responses on NGINX thread pools instead of the event loop. In tests, p99 latency for small requests stayed at 1.7–2.0 ms versus 4.2–198 ms with tokers/zstd-nginx-module, while throughput rose 2.2–15x.

• Small-request p99: 1.7–2.0 ms vs 198 ms for tokers at level 19
• With 8 large requests, tokers latency hit 1.5 s; ngx_mzstd stayed at 2.0 ms
• Throughput 2.2–15x higher on 85 KB–2.2 MB files
• 4m segments cut CPU use to 1.04–1.21x of the event loop