Apply backpressure before reading more frames

[Streetblock]

Summary

Decode complete frames already present in FrameStream's receive buffer before polling the underlying QUIC transport for another chunk.

This prevents a slow frame consumer from causing BufRecvStream to accumulate a new transport chunk on every wake while complete frames are still buffered.

Root cause

FrameStream::poll_next called try_recv before attempting to decode the existing buffer. If one transport chunk contained multiple HTTP/3 frames, every subsequent frame poll could read and append another chunk before consuming the frames already available.

Regression test

The new deterministic test models 64 immediately available transport chunks, each containing 16 HEADERS frames with a 1,024-byte payload, while the consumer processes one frame per wake.

On current master, the test observed:

• 64 underlying transport polls
• 985,920 bytes left buffered after consuming 64 frames
• one encoded transport chunk is 16,432 bytes

With this change:

• only 4 transport polls are needed for the same 64 frames
• buffering remains bounded to the currently consumed transport chunk

Validation

• cargo fmt --all -- --check
• cargo test -p h3 -p h3-quinn
• 228 unit tests passed
• 9 documentation tests passed

Prior and related work

After opening this draft, I noticed the issue timeline links to 0x676e67/http3#2. That June 2026 fork patch already contains the equivalent FrameStream reordering and a smaller two-frame regression test, alongside unrelated QPACK changes. Credit to @0x676e67 for that earlier implementation.

This draft was developed independently against current upstream master; its additional contribution is the stronger multi-chunk regression scenario and measured baseline above. Maintainers may prefer to adapt the earlier commit or this focused patch, and I am happy to adjust attribution or approach accordingly.

The same regression scenario also passes on the broader receive-buffer refactor in #328. This draft offers a smaller near-term change while that larger design is evaluated.

Fixes #308

[seanmonstar]

This looks good to me, reason it's still a draft?

[Streetblock]

Oops, my bad! Leaving it as a draft was just an oversight. I just marked it as ready for review. Thanks!