Test Details Overview

Introduce rpc_host and run every RPC owner on it (#5173)

* Introduce `rpc_host` and run every owner on it

Every owner composed the same three things by hand, an HTTP listener,
the backend that serves its requests and the IO threads they run on,
and each spelled out its own order for stopping them; the test harness
copied the daemon's order with nothing pinning the two together.

`rpc_host` owns the composition. `start` binds the listener and runs
the backend on the host's own IO threads, `stop` closes the listener,
stops the threads and releases the backend, once. `nano_rpc`, the
daemon, the wallet, the RPC test harness and the slow test wrapper run
on it, so in-process RPC now serves on threads sized by the existing
`[process] io_threads` setting instead of the node's.

`rpc_server::stop` becomes idempotent, since on host threads a handler
closing the listener from an IO thread runs concurrently with the
owner's stop; the bound port is cached at start so it stays readable
after the acceptor is closed, and a cancelled accept is no longer
logged as an error. A `bind_failure` test that Windows cannot satisfy
skips itself through new `consteval` platform predicates beside the
sanitizer ones.

* Guard the RPC acceptor against a close racing its accept handler

The accept handler re-arms the acceptor on an IO thread and checked
`is_open` on it, while the owner closes the acceptor from its own
thread. With the listener on the host's own IO threads the two run
concurrently, which ThreadSanitizer reported in
`rpc_host.destroy_without_stop`. Both now take the server's mutex, and
the handler consults the stopped flag instead of the acceptor.

__________

Type: commit

Hash: a5aab87c4b66902c8d4a2ea3657e57a8704f1518

Created: yesterday

__________

gr0v1ty/nano-node:a5aab87c4b66902c8d4a2ea3657e57a8704f1518