[pwolnows]

Description

This PR improves Accuracy Checker validation reliability in two areas:

• selects classification outputs explicitly when a model exposes multiple outputs;
• makes filesystem-derived annotation ordering deterministic before annotation or subset selection.

Changes

Classification adapter

• Prefer an explicitly configured output name when resolving classification model outputs.
• Preserve the existing fallback behavior for models without an explicit output selection.

Deterministic annotation conversion

• Sort image inputs in ImageProcessingConverter, including recursive inputs.
• Sort generated reference images in ParametricImageProcessing before creating Denoise annotations.
• Sort LibriSpeech transcript files before creating annotations.
• Sort CamVid-32 label files before its val_subset_ratio logic selects the validation subset.
• Warn when a parametric image-processing input has no generated reference pair.

Sorting removes dependence on filesystem directory-entry order. Consequently, the same dataset contents produce the same annotation order and the same fixed-seed subset after a dataset restore or annotation-cache regeneration.

Affected Validation Workloads

The deterministic ordering change covers the currently observed unstable validation subsets:

• Denoise
• Colorization
• QuartzNet 15x5 English
• Tiramisu FC-DenseNet-103 / CamVid-32

Validation

• Ran python3 -m compileall for the modified annotation converter modules.
• Ran git diff --check.
• Verified all filesystem enumeration paths touched by the four affected converters are explicitly sorted.
• Checked editor diagnostics for modified converter modules: no errors.

Follow-up

Regenerate validation references for the affected workloads using the deterministic annotation order. Additional annotation converters identified by the broader audit will be addressed separately.