Give up on a hanging vendor before the device gives up on us

[matuszeg]

Closes #29.

upstream_session() created its session with no ClientTimeout, so it
inherited aiohttp's five-minute default. Fine while this only relayed. Not fine
since #21, where a token request is relayed first and answered locally only if
that relay fails — a vendor that hangs rather than refuses would hold the
minted answer far past the point where the device has stopped listening.

Two corrections to the issue, both from measurement

The device's patience is 10 seconds, not 11. The shim log shows
device_got=499 at 10.003 s — the device closing the request. 11 s is its
retry interval; 10 s is where it stops listening.

The bigger problem was not the wait. asyncio.TimeoutError is not an
aiohttp.ClientError, so a timeout escaped _relay's handlers entirely. Three
consequences at once, and the middle one is the worst:

• the device got aiohttp's own 504 instead of the honest 502,
• the message was never parsed, losing a reading already in hand — which is the
  whole point of running this locally,
• and no vendor failure was recorded, so the outage shape that most needs a
  locally minted token would never have opened the gate to get one.

Adding the timeout without catching it would have made that reachable rather
than theoretical.

What shipped

• VENDOR_TIMEOUT_SECONDS = 4 on the session — ~3× the slowest healthy vendor
  response measured at the shim (0.1–1.3 s), and two attempts still fit inside
  the device's 10 s.
• _relay catches TimeoutError beside ClientError, with a written-out
  reason since a bare TimeoutError stringifies to nothing.
• A hang is not retried. The retry exists for the vendor's fast
  FIN-with-no-response, where replaying costs milliseconds; replaying a hang
  doubles the only thing that matters. This falls out of catching
  ServerDisconnectedError alone, and is now pinned by a test rather than left
  implicit.

Tests

A hanging_upstream fixture — accepts the request, answers nothing, holds the
connection — since the existing flaky_upstream models the hangup, which is
a different failure. Four new tests, each verified against the unfixed code:
the two hang tests do not merely fail without the timeout, they sit for five
minutes, which is the defect stated as an experiment.

213 pass. core/ still imports no homeassistant.