fix(detection): don't default tiny80.sod to face-only architecture

[davlaw]

Unrecognized .sod filenames fell through to the :face architecture
fallback, so the tiny80 COCO-80 object model was silently running as
a face-only detector. Adds an exact-match case routing tiny80.sod to
:tiny80, mirroring the existing :voc exact-match case.

Test plan

• Confirmed via journalctl -u lightnvr that tiny80.sod now logs
  Detected COCO-80 model by exact filename match, using :tiny80 architecture
  instead of falling back to face detection, across all streams using this model.

[matteius]

Hey, @davlaw which sod model(s) do you have/use?

[davlaw]

I'm using SOD's Tiny COCO model (tiny80.sod, the 80-class COCO variant, ~65MB) from pixlab.io. This PR came out of debugging why it was silently falling back to the face-only :face architecture — load_sod_model() had an exact-match case for :voc but nothing for tiny80.sod, so it hit the generic unrecognized-.sod-file fallback instead of loading it as :tiny80.