Fix OTA retry accounting and radio delivery races

Keep catalog, manifest and leaf requests pending when the local transmit
queue rejects them instead of spending their network retry allowance.

Keep the packet selected before CAD so work becoming due during sensing
cannot bypass its own channel or pacing checks. Retire cancelled radio
retries before waiting for BUSY hardware or their retry deadline.

Wait for active transmission completion before applying firmware. Stop
admitting new transmissions after the apply drain deadline so busy queues
cannot postpone reboot indefinitely, and keep those deadlines armed across
millis rollover. Retain the existing persistence-flush backoff.

Add failing-before/fixed-after regression coverage for all four issues,
including the production OTA apply and TX guards in CI.

Validation: 1,496 native tests, 42 Python integration tests, and five
firmware builds passed across ESP32-S3, nRF52, and STM32 targets.

docs/ota_user_guide.md

@@ -419,6 +419,9 @@ OTA on either radio profile. `set ota.speed 1` restores the previous pace.
 Ordinary chat, node adverts, and USB/WiFi source transfers keep their timing.
 Local verification of a staged image and leaf comparisons keep their normal
 maintenance rate even when LoRa pacing is slow.
+If the local transmit queue cannot accept a catalog, manifest, or leaf request,
+OTA keeps it pending without consuming a network retry. Successfully queued
+requests that receive no answer still have bounded retries.
 
 `ota config speed 0.5` is an equivalent setter. `ota config speed` and
 `ota speed` read the current factor; the `ota config` summary also includes it.