util/ipc: Add AsyncChannelTransport, the kernel side of the service seam

[chrysh]

The client seam from #464 made real: "every cross-process request goes through ServiceCall, and no process ever blocks on another."

AsyncTransaction lends the kernel 'static buffers for the life of a transaction, which a caller holding an ordinary &[u8] request cannot satisfy. AsyncChannelTransport owns that pair, copies each request in and each response out, and implements util_service::AsyncTransport. One type, every service: register as_raw() with a WaitGroup, start a request, poll once per wake.

Buffer sizes are the wiring's choice. A request longer than the send buffer, or a response longer than the caller's, is TooLarge rather than a truncated frame. Every failure path re-stows the buffers the transaction hands back, so the transport returns to idle and the next call is start.

One change to the layer below, in the first commit: AsyncTransaction::start now takes the request length. The kernel transmitted the whole send buffer, so a transport sized for its largest request also sent the slack behind a short one, and a transport that owns one &'static mut buffer cannot express the subslice any other way.

The QEMU test covers the round-trip, a request that does not fit, cancel then reuse, and the not-ready-then-ready sequence. That last one is the case #484 called out as untestable against a loopback, whose response is always ready on the first poll.

Not in here: the generic Loopback<D: Dispatch>, the blocking ChannelTransport (nothing on the #464 path blocks), and the typed PLDM client. Those follow, then the i2c and pldm migrations onto the shared seam.

Review the last commit only; below it are #482 and #485.

Progress toward 9elements#13.