Feat/platform agnostic tcp connectors

[lxsaah]

Wave B of design 052, first connector: the TCP connector moves onto the runtime-neutral I/O layer wave A added. The adapter owns the socket; this crate contributes only length-prefix framing.

The crate goes from 1 428 lines to 385.

Commits

add	generic TcpClient/TcpServer over StreamDialer/StreamListener; the runtime modules stay
migrate	tests and tcp_demo onto the adapter transports; the old modules become unreferenced
delete	tokio_transport.rs (302) + embassy_transport.rs (591), and the dependencies they needed

Split that way so the deletion diff is purely subtractive and every commit is green on its own.

Result

• Zero unsafe impls in any connector crate — the three here were the last outside the adapters (acceptance 2).
• No tokio, embassy-net, embassy-futures or embedded-io-async in the library graph — checked with cargo tree, not just by deleting files (acceptance 3). embassy-net remains test-only for the loopback harness.

Breaking

lib.rs's whole all(tokio-runtime, embassy-runtime) aliasing block is gone — TokioTcpServer/EmbassyTcpServer, TokioTcpConnection/EmbassyTcpConnection and friends existed only to disambiguate two implementations of the same concept. One implementation, one name:

pub use connector::{framed_dialer, framed_dialer_at, framed_listener, split_host_port,
                    TcpClient, TcpServer, DEFAULT_PORT};

TcpServer::new now takes an already-bound listener rather than a bind string, so the adapter does the binding.

Worth a look in review

Framer::encode is infallible, but length-prefix encoding is not. An oversized frame is dropped whole rather than written half-encoded — a length prefix with no payload behind it would desync the peer permanently, and unlike COBS there is no delimiter to resync on.

TcpServer holds its listener in a OneShot, taken on first poll rather than at call time. A build() future dropped before being polled must leave the listener in place, or a later build fails having never served anything. an_unpolled_build_leaves_the_listener_in_place pins it.

split_host_port fixes a bug it inherited. Two call sites had grown private copies using rsplit_once(':'), which mangles a bracketed IPv6 host with no port — the form aimdb-client's own error message documents. Hoisted into the connector with five tests; brackets are now stripped, which is what both adapters actually resolve.

Verification

make check clean: 2 031 tests, 0 failures, all stages. aimdb-client needed a fix — it used the deleted TcpDialer through the crate-root re-export, which a module-path grep does not see.