Add bidirectional attention support to Gemma 3

[dkuku]

This PR adds support for bidirectional attention in Bumblebee.Text.Gemma3Text.

Motivation

In huggingface/transformers, Gemma3TextConfig introduces the use_bidirectional_attention boolean configuration option. When set to true, the model uses non-causal (bidirectional) attention across all layers rather than causal masking. This is required for embedding and encoder tasks built on top of the Gemma 3 backbone (such as EmbeddingGemma).

Changes

• Added :use_bidirectional_attention option (defaulting to false) to Bumblebee.Text.Gemma3Text.
• Configured causal: not spec.use_bidirectional_attention when building transformer blocks.
• Mapped use_bidirectional_attention in Bumblebee.HuggingFace.Transformers.Config.load/2.
• Added integration test in test/bumblebee/text/gemma3_text_test.exs with reference values generated from Python transformers.

Reference verification (Python)

Output verified against PyTorch transformers with atol: 1.0e-4 using bumblebee-testing/tiny-random-Gemma3TextModel:

import torch
from transformers import AutoModel, AutoConfig

model_id = "bumblebee-testing/tiny-random-Gemma3TextModel"
config = AutoConfig.from_pretrained(model_id)
config.use_bidirectional_attention = True
model = AutoModel.from_pretrained(model_id, config=config)
model.eval()

inputs = {
    "input_ids": torch.tensor([[10, 20, 30, 40, 50, 60, 70, 80, 0, 0]]),
    "attention_mask": torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0]])
}

with torch.no_grad():
    out = model(**inputs)
    print(out.last_hidden_state[:, 1:4, 1:4])
    # => tensor([[[ 0.3755, -0.6418,  1.9320],
    # =>          [-0.5731,  0.3630,  2.5736],
    # =>          [-0.0530,  0.5178,  1.3310]]])

[jonatanklosko]

Please rebase on top of main, and then we can merge this.

[dkuku]

👍🏻 Done