PE not getting updated at runtime when resolution is changed in v1.6.5

### Search before asking - [x] I have searched the RF-DETR issues and found no similar bug report. ### Bug __init__ 2026-04-30T09:28:15.205008636Z     self.rfdetr = self._load_model(weights=self.model_weights, resolution=self.imgsz, device=self.device_by_name, 2026-04-30T09:28:15.205010320Z   File "/app/cvblackjack/packages/common/backbone_architecture/rfdetr.py", line 402, in _load_model 2026-04-30T09:28:15.205011912Z     model = Model(pretrain_weights=weights, 2026-04-30T09:28:15.205013301Z   File "/usr/local/lib/python3.10/dist-packages/rfdetr/detr.py", line 382, in __init__ 2026-04-30T09:28:15.205016450Z     self.model = self.get_model(self.model_config) 2026-04-30T09:28:15.205018033Z   File "/usr/local/lib/python3.10/dist-packages/rfdetr/detr.py", line 1027, in get_model 2026-04-30T09:28:15.205019621Z     return _build_model_context(config) 2026-04-30T09:28:15.205021380Z   File "/usr/local/lib/python3.10/dist-packages/rfdetr/detr.py", line 236, in _build_model_context 2026-04-30T09:28:15.205022988Z     class_names = _load_pretrain_weights_into(nn_model, args) 2026-04-30T09:28:15.205024572Z   File "/usr/local/lib/python3.10/dist-packages/rfdetr/detr.py", line 169, in _load_pretrain_weights_into 2026-04-30T09:28:15.205026147Z     nn_model.load_state_dict(checkpoint["model"], strict=False) 2026-04-30T09:28:15.205027655Z   File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 2629, in load_state_dict 2026-04-30T09:28:15.205029273Z     raise RuntimeError( 2026-04-30T09:28:15.205030706Z RuntimeError: Error(s) in loading state_dict for LWDETR: 2026-04-30T09:28:15.205032695Z  size mismatch for backbone.0.encoder.encoder.embeddings.position_embeddings: copying a param with shape torch.Size([1, 1297, 384]) from checkpoint, the shape in current model is torch.Size([1, 1765, 384]). ### Environment RF-DETR v1.6.5 ### Minimal Reproducible Example Tried running the finetuned rfdetr- seg medium trained on my data with updated resolution of 552 , trained using package v1.6.4 then updated the package passed the custom resolution and got this runtime error of pe mismatch which should have been taken care of internally according to rfdetr v1.6.5 change log ### Additional _No response_ ### Are you willing to submit a PR? - [ ] Yes, I'd like to help by submitting a PR!