Say which silence it is when the dashboard goes quiet

[matuszeg]

A vendor outage and a broken redirect look identical on the dashboard: both end with every chip stale. binary_sensor.pumpspy_local_vendor_reachable has been able to tell them apart since #19, and the dashboard never read it.

Two changes, neither of which takes up space while things are working:

• The liveness chip names the suspect when it goes stale. "No data, vendor down" when the vendor sensor is off, "No data, vendor was OK" when it's on, plain "No data" when nothing has been forwarded yet. The past tense is load-bearing: nothing is forwarded while the device is quiet, so the verdict freezes, and it's the last answer rather than a live one. It's still the useful one, since it was taken at about the moment the reports stopped.
• A card that exists only while the vendor is unreachable, saying the sump is still being watched locally and it's the phone app that's gone deaf. Built-in markdown, not Mushroom, for the same reason as the alerts above it.

No permanent tile, which answers the issue's open question. Vendor reachability is on essentially all the time, and something that reads green for months is something you stop reading.

test_dashboard.py needed two adjustments. The vendor sensor belongs to the integration's own service device, so it has no device id in it and it isn't built from an entity description: the known-id set is rebuilt from descriptions only, and every referenced id had to contain the your_device_id placeholder. Both now allow the two service entities by name, listed one by one rather than exempting anything that lacks a device id, which would exempt a typo too. A new test pins the coupling, checking the raw text as well as the parsed YAML because one of the two references lives inside a Jinja template the entity walk can't see.

Verified on the live install, not just in tests:

• All four branches of the chip template rendered through Home Assistant's own template API, the "Live" branch against real state and the three stale branches with the age forced.
• binary_sensor.pumpspy_local_vendor_reachable confirmed present in /api/states, which is the failure this would otherwise hit silently: a wrong id means the card simply never appears.
• Dashboard deployed and read back identical to what was sent, with every entity it names present on the instance.

One gap, stated plainly: the card itself hasn't been seen on screen, because that needs a real vendor outage and Chrome couldn't reach the instance to screenshot it. Its structure is identical to the three conditional alert cards already running in production.

Closes #28