Add Gemma 4 model family support

Implement the text-only portion of Google DeepMind Gemma 4 architecture:

- Hybrid attention: alternating sliding window and full attention layers
- Dual RoPE: proportional RoPE for full attention, default for sliding
- Per-Layer Embeddings (PLE): per-layer token-dependent gating
- KV sharing: later layers reuse KV from earlier layers of same type
- Q/K/V normalization: RMS normalization on query, key, and value
- Per-layer scalar: learned scaling factor per transformer block
- Optional MoE: mixture-of-experts FFN blocks (26B-A4B variant)

Architectures: :base (Gemma4TextModel), :for_causal_language_modeling
(Gemma4ForCausalLM). Multimodal Gemma4ForConditionalGeneration is not
yet supported.

Uses a custom decoder loop rather than Layers.Transformer.blocks/2
because the model requires features not available in the shared
infrastructure: per-layer embeddings threaded through the block loop,
cross-block KV sharing state, per-layer head dimension variation,
and value normalization.

Includes integration test verified against Python transformers reference
values (atol < 5e-5).

Author: chgeuer

lib/bumblebee.ex

@@ -136,6 +136,9 @@ defmodule Bumblebee do
     "DistilBertForMultipleChoice" => {Bumblebee.Text.Distilbert, :for_multiple_choice},
     "GemmaModel" => {Bumblebee.Text.Gemma, :base},
     "GemmaForCausalLM" => {Bumblebee.Text.Gemma, :for_causal_language_modeling},
+    "Gemma4ForCausalLM" => {Bumblebee.Text.Gemma4, :for_causal_language_modeling},
+    "Gemma4ForConditionalGeneration" => {Bumblebee.Text.Gemma4, :for_causal_language_modeling},
+    "Gemma4TextModel" => {Bumblebee.Text.Gemma4, :base},
     "GemmaForSequenceClassification" => {Bumblebee.Text.Gemma, :for_sequence_classification},
     "Gemma3ForCausalLM" => {Bumblebee.Text.Gemma3Text, :for_causal_language_modeling},
     "Gemma3TextModel" => {Bumblebee.Text.Gemma3Text, :base},
@@ -273,6 +276,8 @@ defmodule Bumblebee do
     "clip" => :clip,
     "gemma" => :gemma,
     "gemma3_text" => :gemma,
+    "gemma4" => :gemma,
+    "gemma4_text" => :gemma,
     "gpt_neox" => :gpt_neo_x,
     "gpt2" => :gpt2,
     "gpt_bigcode" => :gpt2,