Function embed_both 

pub(super) fn embed_both(
    session: &mut Session,
    tokenizer: &Tokenizer,
    texts: &[String],
    cost_model: &CostModel,
    model_variant: ModelVariant,
) -> Result<(Vec<DualEmbedding>, EmbedStats)>

Produces paired dense + sparse embeddings using one session.run() per chunk.

Both projections are derived from the same forward pass:

• FP32: extracts both sentence_embedding (dense) and token_embeddings (sparse base) from the model’s dual outputs.
• FP16/INT8: extracts dense from the CLS token (position 0) of last_hidden_state, and sparse from the full hidden states of the same tensor. This avoids a second forward pass.

Numerically equivalent to calling super::dense::embed_dense and super::sparse::embed_sparse separately, within FP rounding tolerance.