feat: Add Nomic BERT model support

Add support for nomic-ai/nomic-embed-text-v1.5 embedding model.

Architecture:
- Postnorm transformer (standard BERT-style)
- SwiGLU activation (up * silu(gate))
- Rotary position embeddings (RoPE) with base 1000
- Combined Wqkv projection
- No biases in attention and FFN layers
- Mean pooling over non-masked tokens

Tested against Python transformers with ~2e-6 precision.

Author: georgeguimaraes

lib/bumblebee.ex

@@ -176,6 +176,7 @@ defmodule Bumblebee do
     "MPNetForTokenClassification" => {Bumblebee.Text.MpNet, :for_token_classification},
     "MPNetForQuestionAnswering" => {Bumblebee.Text.MpNet, :for_question_answering},
     "MPNetForMultipleChoice" => {Bumblebee.Text.MpNet, :for_multiple_choice},
+    "NomicBertModel" => {Bumblebee.Text.NomicBert, :base},
     "PhiModel" => {Bumblebee.Text.Phi, :base},
     "PhiForCausalLM" => {Bumblebee.Text.Phi, :for_causal_language_modeling},
     "PhiForSequenceClassification" => {Bumblebee.Text.Phi, :for_sequence_classification},
@@ -266,6 +267,7 @@ defmodule Bumblebee do
     "mistral" => :llama,
     "mbart" => :mbart,
     "mpnet" => :mpnet,
+    "nomic_bert" => :bert,
     "phi" => :code_gen,
     "phi3" => :llama,
     "qwen3" => :qwen2,