fix/typing(layer_set): Allow layer set read_proc to return single layer

[nkemnitz]

Allows combining two layers with a read_proc, e.g. https://github.com/ZettaAI/specs/blob/main/nico/inference/chen_s1/11_fine_align_pairwise.cue#L196-L223

Already worked before, just the static type checker would complain.

[nkemnitz]

Tests fail to unrelated zetta_utils/internal/segmentation/embedding/utils.py:117: error: Argument 1 to "_compute_affinity" has incompatible type "ndarray[Any, dtype[Any]]"; expected "Tensor" [arg-type]

[supersergiy]

I think Ideally we would want to handle it in a way that would still keep the return type of layer set's __getitem__ predictable. With this change, every time a user reads from a layer set in python they would have to assert the expected return type. Let me think about this a bit...

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 100.00%. Comparing base (056e9ad) to head (dc38ed6).

Additional details and impacted files

@@            Coverage Diff            @@
##              main      #761   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files          142       142           
  Lines         5877      5877           
=========================================
  Hits          5877      5877           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[nkemnitz]

Any update here?