feat(negotiation): Negotiate through per-device sessions with plain and encrypted outers

[kb1ibt]

Third of four stacked PRs replacing #70; based on #78. It answers the protocol half of your review: each variant is its own class, chosen without a bool, and a device holds the one its firmware speaks.

A device now holds a negotiated session for each connection (SolixBLE/protocols/, typing.Protocol interfaces with @override):

• Outer protocol (PlainOuter / EncryptedOuter): how the negotiation travels and which cipher the session uses. Plain opens with 0001 in the clear and runs CBC; encrypted opens with 4001 under the static GCM key and runs GCM. The session cipher follows the opening frame.
• Path (EcdhPath): the ECDH key exchange that follows the shared opening, on either outer. It sends a fresh P-256 key every negotiation (as Add AS220 (SOLIX S2000) device support #65 and anker-solix-api#345 do), states its method in x005 as the app does (a5 40 on plain, a5 44 with the device's MTU and auth mode on encrypted), and sends x022 with the UTC offset as int32 seconds west and the POSIX time zone.
• Choosing the outer: the advertisement's capability byte when the caller passes it (DeviceClass(ble_device, advertisement=...)), else the outer that last authorized on this instance, else the model's default. A device that drops the link before answering a plain 0001 refuses it, and connect() reopens once with the encrypted outer in the same call; the encrypted outer is never stepped down to plain. DeviceClass(ble_device) with no advertisement keeps working for HaSolixBLE, and type(device) stays the model class.
• Authorization: plain at 0821; encrypted at 4827 status 00, or at the first session push the client can decrypt. A 4827 on the device-initiated pattern 030101 (the grant after a button press) reaches the same handler. A 09 status and its confirmation window are recorded on device.announcement; waiting for the button is left to the pairing PR, which overrides _negotiation_should_restart() / _negotiation_deadline().
• PrimeDevice keeps its UUID, the encrypted default and _post_authorize (4200, 420a); its copy of the negotiation and its own _send_command go.

Prime frames: every Prime command now ends with the typed timestamp trailer fe 05 03 <ts> (the base _send_command), and 420a is built as the app builds it: a2 = region typed 02 (set_region(), else the host locale's, else GB), a3 = this client's identifier as the owner, no a5. These match what the Anker app sends to the Prime Charger 250W (A2345) and the Prime Charging Station (A91B2) and what their firmware reads. I have no 160W or 20K power bank to test on, so their new bytes are untested; could you, or a 160W/20K owner, try it?

Derived, not recorded, test vectors: the C1000 plain 4022 (new layout), the Prime 4200 / 420a, and the 32 Prime command frames with the typed trailer. Each was derived from the recorded session (the pinned test key and the device's recorded public point) and the old frames decrypt to the old plaintexts first. The pinned keys reproduce the public points in the recorded 0021 / 4021, so the recorded negotiations are unchanged up to those frames.

Also: typing_extensions for Python < 3.12 (override); PRIVATE_KEY and NEGOTIATION_COMMAND_* move from const.py to tests/const.py; tests/helpers.py gains install_session_keys, RecordingLink and MockDevice.refuse_after; docs: protocols.rst "Outer protocols".

[kb1ibt]

One follow-up commit on this PR (6d838d2), from running the plain-refusal case against a C2000 Gen 2 with the #81 console:

• A drop during the opening write is a refusal too. The C2000 Gen 2 refuses a plain 0001 by dropping the link, and sometimes the drop lands before the write returns. The write then raised BleakError: disconnected out of connect() instead of reopening with 4001. When the drop came after the write, the reopen already worked: 4801 … 4822 00, 4827 00, connected. Now, if the opening write fails because the link has already dropped, the negotiation treats it as a dropped link, so connect() reopens with the encrypted outer either way. A write error on a link that's still up is still raised.
• MockDevice.refuse_after(during_write=True) drops the link inside the write and raises, as bleak does; the new test failed with the same BleakError before the fix.

#80 and #81 are rebased on it.

Suite: 469 passed (468 before this commit). mypy --strict: 202 errors, unchanged.