fix(proxy): ignore empty body payloads in HTTP/1 upstream writes

[Curious-r]

Summary

Fix a spurious HTTP/1 warning when an empty final body task follows a fully written Content-Length body. The upstream sink now skips empty payloads in both Body and UpgradedBody, while finish_body() still handles final framing and Content-Length validation. Non-empty data after the body is finished continues to trigger the warning.

Added regression tests for Content-Length and chunked bodies to ensure empty final tasks don't affect wire framing or produce duplicate chunk terminators.

Validated with cargo test -p pingora-proxy --lib, cargo clippy -p pingora-proxy --lib --tests -- -D warnings, and cargo fmt --all -- --check.

Background:

When using a Pingora-based reverse proxy for my Matrix homeserver, I noticed that it produced a large number of warnings like:

WARN pingora_proxy::proxy_h1: Upstream body is already finished. Nothing to write

After enabling DEBUG logging, the warning could be reproduced on normal Matrix federation requests. The relevant sequence was:

DEBUG pingora_proxy::proxy_h1: Read 208 bytes body from downstream
DEBUG pingora_proxy::proxy_h1: Read 0 bytes body from downstream
DEBUG pingora_proxy::proxy_h1: Write 208 bytes body to upstream
WARN  pingora_proxy::proxy_h1: Upstream body is already finished. Nothing to write
DEBUG pingora_proxy::proxy_h1: finish sending body to upstream

This shows that the warning is triggered by an empty body event arriving after the complete Content-Length body has already been written. Since empty payloads do not carry any data and finish_body() already handles the end-of-body framing, the empty buffer should be ignored rather than passed to write_body().

[Curious-r]

Sorry, I missed #936 before opening this. Looking at the patches, though, I don't think this is a duplicate — they seem complementary.

#936 hardens BodyWriter at the codec layer by making zero-length writes no-ops in chunked mode, ensuring finish() remains the only place that emits the terminator.

This PR works at the HTTP/1 proxy adapter instead: send_body_to1() skips empty payload buffers before calling the upstream body writer, while still using finish_body() for end-of-stream framing. That covers both Body and UpgradedBody tasks, avoids the spurious "Upstream body is already finished" warning for fully written Content-Length bodies, and includes wire-level regression tests for Content-Length and chunked bodies.

Given the different layers, I think it may make sense to land both: #936 protects the writer API more broadly, while this change keeps the generic proxy task-to-H1 translation from treating an empty EOS signal as application data.