Fix duplicate cross-class form detections

[JasBogans]

Summary

• enable class-agnostic NMS for the default FFDNet inference path
• add a regression test that locks the Ultralytics prediction option in place

Problem

FFDNet currently uses Ultralytics' default class-aware NMS. When the same region
is detected as different field classes, such as a text field and a choice
button, the detections cannot suppress one another and can produce stacked
duplicate PDF fields.

Change

Pass agnostic_nms=True to the default FFDNet prediction call so strongly
overlapping detections compete regardless of predicted class. The fast ONNX
path is unchanged.

Validation

• pytest -q: 8 passed
• ruff check tests/inference_test.py: passed

The five existing Ruff findings in commonforms/inference.py are unchanged by
this PR.

AI assistance disclosure

This change was developed with AI assistance. All changes were reviewed and
validated by the submitter using the test suite.