[Q] Non-causal / bidirectional attention in decoder models for sentence embeddings

[dkuku]

While working on sentence transformers support in Bumblebee, I hit an issue running popular decoder-based text embedding models (such as Alibaba-NLP/gte-Qwen2-7B-instruct or fine-tuned Llama/Mistral embedding backbones). These models require non-causal (bidirectional) attention across input tokens.

In #468, :use_bidirectional_attention (default false) was introduced for Gemma3Text, following Gemma3TextConfig.use_bidirectional_attention in huggingface/transformers.

However, in checkpoints for other decoder architectures (like Qwen, Llama, and Mistral), Hugging Face models typically set "is_causal": false in config.json instead of using use_bidirectional_attention.

Before opening a PR, I'd like to ask how you'd prefer to model this in Bumblebee:

1. Expose :is_causal (default: true):
   Use :is_causal directly on the spec, or accept both and validate against conflicting values in Gemma - https://github.com/huggingface/transformers/blob/469230357aab0f2b303b0d638c1f8d06edb14184/src/transformers/models/gemma3/modeling_gemma3.py#L317.
2. Unify on :use_bidirectional_attention (default: false):
   Add :use_bidirectional_attention to Shared.common_options/1 across Llama, Mistral, Qwen3, and Gemma 3, map "is_causal": false from HF configs, and validate against conflicting values in shared options.