Validate buffered read ownership during Windows TLS upgrades

Decision: Needs information - plausible Windows-specific native lifetime risk.

Evidence: `zig/transport.zig:338-373` can retain a `Py_buffer` while a Windows overlapped read is pending. `pauseReading()` records `READ_STATE_PENDING`, but `setProtocol()` at `zig/transport.zig:927-931` releases the view unconditionally. `start_tls()` pauses reading and immediately replaces the protocol.

Impact: If libuv retains the supplied buffer across loop iterations on Windows, a TLS upgrade under active buffered traffic could release memory while the kernel is still writing to it.

Next action: Add an AMD64 and ARM64 Windows regression using `BufferedProtocol`, continuous reads, and repeated `start_tls()` or protocol replacement under debug allocation checks. Confirm libuv's buffer lifetime for this path.

Acceptance criteria: Either the test proves the current ordering safe and documents why, or the buffer release is deferred until the matching read callback completes.

Read zig/transport.zig:338-373 and 927-931, then trace pauseReading(), setProtocol(), and start_tls() during a pending BufferedProtocol read. Add the AMD64 and ARM64 Windows regression with continuous reads and repeated TLS upgrades or protocol replacement, using debug allocation checks; done means the test documents safe ordering or verifies release waits for the matching read callback.

Written by the indexing model from the issue text.