Fix detector training with AMP enabled

[sopenlaz0]

What changed

Initialize Trainer.use_transformers to False in the detector model setup and add a focused detector-AMP training regression test.

Root cause

The AMP training branch checks self.use_transformers before choosing the model call signature. Recognition setup initializes that attribute, but detector setup did not, so detector training with Global.use_amp: true raised AttributeError before its first forward pass.

Detection models always use the standard model(image, data=...) path, so setting the flag to False in _init_det_model makes that contract explicit.

Validation

• pytest -q tests/test_detector_amp.py — 1 passed
• Structured Codex autoreview — clean, no actionable findings

The regression test initializes the detector builders and executes one AMP optimizer step through Trainer.train; it fails on the previous code at the undefined attribute and passes with this fix.