fix(http1): drain client dispatch receiver on connection shutdown

[akasakariko]

Closes #4202

Summary

On HTTP/1 shutdown, tasks reserving channel capacity before close could publish afterward and leave envelopes stranded; this fix closes and asynchronously drains the dispatch receiver until Ready(None), canceling unsent requests with original requests recoverable while callers hold SendRequest or hyper-util

Changes:

• src/proto/h1/dispatch.rs and src/client/dispatch.rs: close and drain dispatch receiver on shutdown until Ready(None)
• Cache original connection result during drain to avoid polling closed IO again
• Keep upgrade and CONNECT handoff behavior unchanged
• Regression tests use Tokio budget exhaustion for reset, idle EOF, and dispatch driving without IO shutdown
• Stranded callbacks reproduced separately with temporary reservation-splitting instrumentation; instrumentation excluded from commit

Verification

Local verification passed (GitHub CI not started):

• cargo test --features full --lib client_close_drains_pending_requests: 3 pass with fix, 3 fail on original code; passes on Tokio 1.38.2
• cargo test --features full: 315 passed, 10 ignored
• rustfmt --check --edition 2021 $(git ls-files '*.rs') and git diff --check pass
• cargo clippy --features full -- -D warnings: exits 0 with 3 pre-existing manifest warnings
• cargo hack depth-2 powerset passes across 66 configurations including tracing
• Rust 1.63.0 cargo check passes with MSRV-aware vendored dependencies
• cargo semver-checks against origin/master: 223 checks pass
• RUSTFLAGS='--cfg hyper_unstable_ffi' cargo test --features client,http1,http2,ffi --lib: 99 passed, 6 ignored

[seanmonstar]

Thanks!