Replace stdlib gzip/flate with klauspost/compress in compress filter

[ivan-digital]

Summary

Replace compress/gzip and compress/flate with github.com/klauspost/compress in the compress filter. Drop-in API-compatible replacement with significantly better CPU performance.

Scoped to the compress filter only — decompress, routesrv, and OIDC cookie compression can follow up separately.

Motivation

Profiling shows 30% CPU spent in compress/flate on high-traffic clusters (#3932).

Benchmark results (Apple M2 Max, 6 samples)

Default level (BestSpeed — production path)

                     │  stdlib   │ klauspost │  change  │
CompressGzip4  10KB    16.12µs    9.38µs      -42%
CompressGzip6  1MB     985.9µs   401.1µs      -59%
CompressGzip8  100MB   88.76ms   24.56ms      -72%

JSON payload (compressible, realistic)

                     │  stdlib   │ klauspost │  change  │
GzipJSON6      1MB    113.9µs    72.4µs      -36%
GzipJSON8      100MB  10.68ms    6.77ms      -37%

Deflate-only

                     │  stdlib   │ klauspost │  change  │
Deflate4       10KB   16.50µs    9.06µs      -45%
Deflate6       1MB    1041.2µs   366.0µs     -65%

Higher compression level (level 6)

                     │  stdlib   │ klauspost │  change  │
GzipLevel6_6   1MB   2293.3µs   464.0µs     -80%

Memory allocation nearly identical. Brotli unchanged (separate library).

Closes #3932

Test plan

• All existing compress filter tests pass
• Benchmarks across random data, JSON, deflate, and compression levels

[szuecs]

thanks for the PR, I took the test to increase coverage on my PR.