fix: reimplement BertModelWarper utility methods for transformers v5 compatibility

[djdarcy]

TLDR version:

• Fixes GroundingDINO model loading crash with transformers >= 5.0
• Reimplements three utility methods (get_head_mask, get_extended_attention_mask, invert_attention_mask) as local class methods in BertModelWarper instead of copying them as bound references from the BertModel instance

Problem

BertModelWarper.__init__ previously copied three methods directly from the BertModel instance:

self.get_extended_attention_mask = bert_model.get_extended_attention_mask
self.invert_attention_mask = bert_model.invert_attention_mask
self.get_head_mask = bert_model.get_head_mask

Transformers v5.0 broke this pattern in multiple ways:

1. get_head_mask was removed entirely (huggingface/transformers PR #41076), causing AttributeError: 'BertModel' object has no attribute 'get_head_mask'
2. get_extended_attention_mask changed its 3rd parameter from device to dtype, causing TypeError: to() received an invalid combination of arguments - got (dtype=torch.device, )
3. invert_attention_mask still works today but uses the same fragile bound-method pattern and references self.dtype internally

Fix

All three methods are reimplemented as proper class methods on BertModelWarper. The logic is identical to the original transformers implementations -- these are simple mask reshaping and dtype casting operations with no model-state dependencies beyond self.config.

This makes BertModelWarper fully self-contained and independent of the transformers library version.

Some basic testing done

• Tested with transformers==5.3.0 -- GroundingDINO model loads and runs inference successfully
• The reimplemented methods produce the same mask shapes and values as the originals

Closes #110
Closes #111

[AmrKhaledDar]

It works thx a lot

[magiak]

Confirming this PR fixes the 'BertModel' object has no attribute 'get_head_mask' error for me.

Environment:

ComfyUI 0.3.40
Python 3.11.10
torch 2.5.1+cu124
NVIDIA RTX 4090
Checked out the PR branch (git fetch origin pull/113/head:fix-113 && git checkout fix-113), restarted ComfyUI, and GroundingDinoSAMSegment now loads weights and produces correct masks from text prompts. Tested with SAM sam_vit_b + GroundingDINO_SwinT_OGC. 👍

Would be great to see this merged.