Fetch each port once and drop the fixed 10 second refresh timeout

[tkolsto]

Problem

ACInfinityDataUpdateCoordinator._async_update_data wraps the entire refresh in a fixed async_timeout.timeout(10). But refresh() makes one sequential API call per controller plus per port — and it currently fetches every port twice, since port controls and port settings both come from the same getdevModeSettingList response.

Refresh duration therefore grows linearly with account size. Measured against a real account with 13 controllers / 68 ports: 150 sequential calls at ~288 ms/call with connection reuse ≈ 43 seconds — against a 10 second ceiling. Once an account grows past roughly 3 controllers' worth of ports, no refresh can ever complete: every poll times out, UpdateFailed is raised, and every entity of every controller is permanently unavailable. Removing and re-adding the integration doesn't help, because the first refresh fails the same way.

The user-visible symptom is "the integration just stopped getting data one day" — the day the account crossed the size threshold. Field-confirmed: a user with exactly this symptom (all devices visible, every value blank, reinstall didn't help) had data restored immediately by this change.

This may also explain the anomalous ~1m49s polling gap in the debug log attached to #146.

Changes

1. Fetch each port once. Store both port controls and port settings from a single getdevModeSettingList response instead of requesting it twice. Halves the number of calls per refresh.
2. Remove the overall 10 second refresh timeout. Every request already has its own 10 second timeout inside the client, and the retry logic bounds failures, so total refresh duration is already bounded — the blanket wrapper adds nothing except a hard ceiling on account size.

For users with larger accounts it's still worth raising the polling interval in the integration options, since a refresh that takes ~20 s shouldn't be scheduled every 10 s — but with this change a large account degrades to "slow" instead of "permanently unavailable". (Parallelizing the per-port fetches would speed this up further, but given the API's sensitivity to request bursts (#39) I kept this PR to the conservative fixes.)

Testing

• Full suite passes: 1764 (1763 baseline + 1 new regression test asserting exactly one settings fetch per port, with controls and settings both populated from it). pyright clean.
• Field-verified on a 13-controller account that was fully dead on 2.2.0: data returned immediately after installing a build with this change.

Independent of #149 and #150; merges cleanly with both.