feat: Add ModernBERT model family

Add support for ModernBERT, a modern encoder model with architectural improvements over BERT:

- Rotary position embeddings (RoPE) instead of absolute position embeddings
- Alternating local and global attention layers for efficiency
- Gated linear units (GeGLU) in feed-forward blocks
- Pre-normalization with LayerNorm (no bias)
- First layer reuses embedding norm for attention

Supported architectures:
- :base
- :for_masked_language_modeling
- :for_sequence_classification
- :for_token_classification

Reference: https://arxiv.org/abs/2412.13663

Author: georgeguimaraes

lib/bumblebee.ex

@@ -170,6 +170,11 @@ defmodule Bumblebee do
     "MistralModel" => {Bumblebee.Text.Mistral, :base},
     "MistralForCausalLM" => {Bumblebee.Text.Mistral, :for_causal_language_modeling},
     "MistralForSequenceClassification" => {Bumblebee.Text.Mistral, :for_sequence_classification},
+    "ModernBertModel" => {Bumblebee.Text.ModernBert, :base},
+    "ModernBertForMaskedLM" => {Bumblebee.Text.ModernBert, :for_masked_language_modeling},
+    "ModernBertForSequenceClassification" =>
+      {Bumblebee.Text.ModernBert, :for_sequence_classification},
+    "ModernBertForTokenClassification" => {Bumblebee.Text.ModernBert, :for_token_classification},
     "PhiModel" => {Bumblebee.Text.Phi, :base},
     "PhiForCausalLM" => {Bumblebee.Text.Phi, :for_causal_language_modeling},
     "PhiForSequenceClassification" => {Bumblebee.Text.Phi, :for_sequence_classification},
@@ -259,6 +264,7 @@ defmodule Bumblebee do
     "llama" => :llama,
     "mistral" => :llama,
     "mbart" => :mbart,
+    "modernbert" => :modernbert,
     "phi" => :code_gen,
     "phi3" => :llama,
     "qwen3" => :qwen2,