[model] Support gliner2-multi-v1 model

[Seunghui98]

What

Let's support the gliner2-multi-v2 model for Named Entity Recognition (NER).

• This model utilizes microsoft/mdeberta-v3-base as its back-bone.
• It introduces a specialized architecture (GLiNER) for span representation and entity counting.

How

We need to implement the DebertaV2 architecture and GLiNER-specific modules.

1. DebertaV2 Backbone

• The core difference from standard BERT is the Disentangled Attention mechanism.
• We need to implement DebertaAttentionLayer which handles Content-to-Content(c2c), Content-to-Position(c2p), and Position-to-Content(p2c) attention scores separately.
• Relative positional embeddings shoud be handled dynamically.

2. GLiNER Specialized Layers

• SpanRepLayer : Aggregates token embeddings into span repreentations by concatenating start and end token projection.
• CountLSTMLayer : Predicts the number of entities in the text using a sequence modeling approach.

TODO

• DebertV2 Backbone encoder Implementation
  • Implement DebertaLayer and DebertaV2 model]([causallm] Support deberta_v2 embedding model nntrainer#3773)
• GLiNER Specific modules
  • Implement SpanRepLayer
  • Implemnet CountLSTMLayer

FYI

• hugging face - gliner2-multi-v1
• transformers - deberta_v2

model structure

============================================================
🧠 Model Configuration
============================================================
Encoder model      : microsoft/mdeberta-v3-base
Counting layer     : count_lstm
Token pooling      : first
============================================================

GLiNER2(
  (encoder): DebertaV2Model(
    (embeddings): DebertaV2Embeddings(
      (word_embeddings): Embedding(250112, 768, padding_idx=0)
      (LayerNorm): LayerNorm((768,), eps=1e-07, elementwise_affine=True)
      (dropout): Dropout(p=0.1, inplace=False)
    )
    (encoder): DebertaV2Encoder(
      (layer): ModuleList(
        (0-11): 12 x DebertaV2Layer(
          (attention): DebertaV2Attention(
            (self): DisentangledSelfAttention(
              (query_proj): Linear(in_features=768, out_features=768, bias=True)
              (key_proj): Linear(in_features=768, out_features=768, bias=True)
              (value_proj): Linear(in_features=768, out_features=768, bias=True)
              (pos_dropout): Dropout(p=0.1, inplace=False)
              (dropout): Dropout(p=0.1, inplace=False)
            )
            (output): DebertaV2SelfOutput(
              (dense): Linear(in_features=768, out_features=768, bias=True)
              (LayerNorm): LayerNorm((768,), eps=1e-07, elementwise_affine=True)
              (dropout): Dropout(p=0.1, inplace=False)
            )
          )
          (intermediate): DebertaV2Intermediate(
            (dense): Linear(in_features=768, out_features=3072, bias=True)
            (intermediate_act_fn): GELUActivation()
          )
          (output): DebertaV2Output(
            (dense): Linear(in_features=3072, out_features=768, bias=True)
            (LayerNorm): LayerNorm((768,), eps=1e-07, elementwise_affine=True)
            (dropout): Dropout(p=0.1, inplace=False)
          )
        )
      )
      (rel_embeddings): Embedding(512, 768)
      (LayerNorm): LayerNorm((768,), eps=1e-07, elementwise_affine=True)
    )
  )

  (span_rep): SpanRepLayer(
    (span_rep_layer): SpanMarkerV0(
      (project_start): Sequential(
        (0): Linear(in_features=768, out_features=3072, bias=True)
        (1): ReLU()
        (2): Dropout(p=0.1, inplace=False)
        (3): Linear(in_features=3072, out_features=768, bias=True)
      )
      (project_end): Sequential(
        (0): Linear(in_features=768, out_features=3072, bias=True)
        (1): ReLU()
        (2): Dropout(p=0.1, inplace=False)
        (3): Linear(in_features=3072, out_features=768, bias=True)
      )
      (out_project): Sequential(
        (0): Linear(in_features=1536, out_features=3072, bias=True)
        (1): ReLU()
        (2): Dropout(p=0.1, inplace=False)
        (3): Linear(in_features=3072, out_features=768, bias=True)
      )
    )
  )
  (classifier): Sequential(
    (0): Linear(in_features=768, out_features=1536, bias=True)
    (1): ReLU()
    (2): Linear(in_features=1536, out_features=1, bias=True)
  )
  (count_pred): Sequential(
    (0): Linear(in_features=768, out_features=1536, bias=True)
    (1): ReLU()
    (2): Linear(in_features=1536, out_features=20, bias=True)
  )
  (count_embed): CountLSTM(
    (pos_embedding): Embedding(20, 768)
    (gru): GRU(768, 768)
    (projector): Sequential(
      (0): Linear(in_features=1536, out_features=3072, bias=True)
      (1): ReLU()
      (2): Linear(in_features=3072, out_features=768, bias=True)
    )
  )
)