perf: use klauspost/compress for gzip and flate decompression

[burruplambert]

Drop-in replacement of compress/gzip and compress/flate with github.com/klauspost/compress for response body decompression.

The klauspost implementation has a significantly faster Huffman decoder. In CPU profiling of a high-throughput production client workload, stdlib gzip decompression accounted for roughly 5% of total CPU, and swapping the import removed most of that.

• No API or wire behavior change - the klauspost gzip and flate packages are drop-in compatible with the stdlib ones
• No new dependency - the module already requires klauspost/compress
• Supersedes the decompression half of perf: reduce CPU overhead in gzip decompression and header lowercasing #17, which bundled two unrelated changes and is being closed in favor of focused PRs

Note: the package test suite reports a failure at the end of any full run because TestMain's goroutine leak checker flags its own goroutine. That is pre-existing on master and unrelated to this change.

[burruplambert]

Checked the drop-in claim rather than relying on it. Compared klauspost against stdlib across 14 malformed-input cases - empty input, non-gzip data, truncated header, truncated body, truncated trailer, bad magic, bad compression method, corrupt CRC, corrupt size, corrupt deflate stream, trailing garbage, and concatenated streams, plus a corrupt raw flate stream.

Every case produced the identical outcome, including the exact error text - gzip: invalid header, gzip: invalid checksum, unexpected EOF, and even flate: corrupt input before offset 25. No case differed, and neither implementation panicked where the other did not.

So callers that inspect or match on decompression errors see no change.