fix: re-handshake in place so a dropped session costs 0 polls, not 1

[agjs]

The bug

Every plug was answering only about half its polls in production:

avg_over_time(tapo_plug_up[3h])
  ASUS Ascent GX10  0.503
  DGX Spark         0.503
  k3s Cluster       0.503

WARNING Response error: 403 Forbidden
WARNING poll failed for plug DGX Spark (192.168.10.217): Tapo(SessionTimeout)
WARNING poll failed for plug ASUS Ascent GX10 (192.168.10.216): Tapo(SessionTimeout)

Three plugs, three IPs, identical 0.503 — that is not flaky hardware. The alternation was built into poll().

A Tapo session here survives exactly one full read; the next request answers 403 Forbidden. The old code reacted by setting up=0, clearing self.device, and waiting for the next cycle to handshake. So the loop could only ever go: success → expiry → re-handshake → expiry → forever.

The fix

poll() retries once in place: on failure it drops the session, re-handshakes, and reads again — reporting up=0 only if that also fails. A recoverable expiry now costs zero polls.

Each poller also gets its own ApiClient. Sharing one across plugs let their handshakes race, which is the likeliest reason a session was being invalidated after a single use to begin with.

Verified

test_poller.py models the plug that was actually observed — a session that dies after one complete read — and runs the real PlugPoller:

old code:  up per poll  : [1, 0, 1, 0, 1, 0, 1, 0, 1, 0]     success rate: 50%
new code:  up per poll  : [1, 1, 1, 1, 1, 1, 1, 1, 1, 1]     success rate: 100%
           reauths: 9    hard failures: 0

A second case checks the retry does not paper over a plug that is genuinely gone — an unreachable plug still reports up=0 and increments the failure counter.

Three new metrics

This failure was near-invisible from outside: the process stayed up, Prometheus scraped it happily, and the power gauges kept moving from the successful half of the polls. Readings looked live — just arriving at half the intended rate, with energy totals quietly drifting.

Metric	Why
tapo_plug_last_success_seconds	gauges keep their last good values on failure, so nothing ever looks missing. time() - this is the true age of every reading.
tapo_plug_reauth_total	recoverable expiries, absorbed silently — a rising rate means the session lifetime is degrading
tapo_plug_poll_failures_total	polls that failed even after a re-handshake

CI

The test runs on pull requests as well as pushes; :latest still publishes only from main.

Consumer side

Alerting added separately in argocd-app-of-apps (manifests/observability/monitoring/tapo-alerts.yaml) so a half-dead exporter cannot hide again — TapoPlugPollsFailing fires below 90% poll success over 30m.

Closes #1