Fix the issues found in the review of 1.0.4

[DAB-LABS]

I had 1.0.4 reviewed a fifth time. The library came through clean, the review found one more way the new socket could report an error the original never did.

The socket change is small. When a device is off, some routers answer for it with a "host unreachable" message. The original library never saw those and simply timed out, and Home Assistant gives a timeout a few tries before marking a device unavailable, but marks it unavailable at once on a socket error. So those messages are now treated as silence, the way the original behaved. I measured this on the bench first: on my network neither case actually produced the message, so this is insurance for networks that do.

The capture change replaces the mechanism that decides whether a capture window is free with a plain flag. Same behaviour, same tests. Tested on Python 3.13 and 3.14, 272 tests, plus a live run against an RM4 Pro.

Coding assistant's notes on this change:

EHOSTUNREACH and Windows's ConnectionResetError join ECONNREFUSED in the "silence" branch of _exchange: logged, the timeout decides, the socket is dropped afterwards so the next call heals. Bench measurement before the change: an on-link address with no host behind it (ARP incomplete) and an off-subnet address both produced NetworkTimeoutError after the full timeout on the test network, with no ICMP surfacing.

The capture window claim is now a flag set on the window's first iteration and cleared in the generator's finally, which asyncio's finalizer runs when a dropped generator is collected; the weak reference and ag_frame inspection are gone. The existing tests (dropped window does not block, held window is refused, paused consumer keeps its window, unstarted window does not block) pin the behaviour unchanged. One visible difference: CaptureInProgressError always comes from the new window's first iteration now, never from the capture() call itself. Also: discover() closes the xdiscover() generator it drains; a comment explains why the RM Max sits in rmpro (upstream mjg59#838's text says rm4pro, its tested diff says rmpro); the README notes that a device belongs to the event loop it first talks on. Typing, py.typed and the CLI move are deliberately left for a later release.