Support batched embeddings

[iamlemec]

This allows for efficient high volume embedding. Changes include:

• Final pooling layer now sums by sequence id to compute correct embedding for batches with multiples sequences
• Pooling layer can be toggled with do_pooling. Default is true but false may be useful for ColBERT approaches
• Bring back non-causal attention mask, which got lost in the creation of llama_set_inputs going to alloc v3
• Embeddings can accessed individually by seq_id with llama_get_embeddings_ith
• Updated embedding example to split by newline and group into batches by default

Performance looks to be on par with ONNX for CPU/GPU, at least for relatively large models such as bge-base where tokenization is not a bottleneck.