perf(client) + fix(h1): decode/body hot-path optimizations and a single-connection HTTP/1.1 pipelining deadlock

[st0o0]

Overview

Five client hot-path performance optimizations on the HTTP/1.1 and HTTP/2 decode/body
paths, plus a fix for an intermittent HTTP/1.1 single-connection pipelining deadlock
that surfaced while stress-testing them. All behavior-preserving except one documented
internal decoder contract change.

Fix - HTTP/1.1 single-connection pipelining deadlock

Under heavy single-connection pipelining the server streams responses back to back, so a
response's status line or header block is frequently split across two TCP reads. The
H1.1 client kept no cross-read remainder (only the streaming back-pressure path retained
anything), so the unconsumed prefix of a split header was discarded and the next read's
continuation parsed as garbage → HttpProtocolException: Malformed header field. That
faulted one request and stranded its in-flight pipelined siblings, deadlocking the
connection (SingleConnectionConcurrencyRegressionSpec failed ~50–80% of runs).

Fix: retain the unconsumed prefix in Http11ClientStateMachine (_partialResponse)
and prepend it to the next inbound buffer - mirroring what the H2 FrameDecoder already
does - cleared on disconnect, decode failure, and cleanup. Added a deterministic repro
(Http11ClientFragmentedResponseSpec) covering status-line, header-line, and
second-pipelined-response splits. The stress guard now drains 256 × 40 cleanly.

Performance optimizations

• H2 FrameDecoder.Decode returns its reused _frames list directly instead of
  ToArray() per inbound read — removes one array allocation per H2 network read on both
  client and server. Callers that hold a result across a later Decode now snapshot
  explicitly; CLAUDE.md updated to document the contract.
• HPACK encoder/decoder reuse the raw UTF-8 byte lengths they already computed instead
  of recomputing GetByteCount when adding to the dynamic table (single 4-arg Add).
• H1.x BufferSearch.FindCrlf uses a single vectorized two-byte IndexOf("\r\n"u8)
  instead of find-CR-then-check-LF with scan restarts — runs once per response line.
• QueuedBodyStream.CopyToAsync override writes pooled body chunks straight to the
  destination, removing a body-sized copy and the 81,920-byte framework rent per buffered
  download (return-to-pool deferred until after the write completes).

Testing

All test projects green, no flakes:
unit 5754/0 · End2End 101/0 · Client 474/0 · Server 89/0 · AcceptanceTests 712/0 ·
API.Tests 1/0 · AspNetCore.Tests 36/0. Each change is TDD'd with a deterministic spec;
Roslyn diagnostics clean on every changed file.

Notes

• Submodule lib/servus.akka is unchanged (gitlink not bumped).
• The only behavioral contract change is FrameDecoder.Decode returning a reused buffer;
  all in-tree callers consume it synchronously and tests that held results were updated to
  snapshot.