Support num_key_value_heads configuration for grouped-query attention.

[jnp]

This patch fixes the issue described in #4027

In summary the llama2 model uses grouped query attention therefore it has a new configuration for key value heads in addition to num_heads.
The configurations for llama2-13b-chat model is as follows:
{
"_name_or_path": null,
"architectures": [
"LlamaForCausalLM"
],
"bos_token_id": 1,
"eos_token_id": 2,
"hidden_act": "silu",
"hidden_size": 5120,
"initializer_range": 0.02,
"intermediate_size": 13824,
"max_position_embeddings": 2048,
"model_type": "llama",
"num_attention_heads": 40,
"num_hidden_layers": 40,
"num_key_value_heads": 40,
"pad_token_id": 0,
"pretraining_tp": 1,
"rms_norm_eps": 1e-05,
"rope_scaling": null,
"tie_word_embeddings": false,
"torch_dtype": "float16",
"transformers_version": "4.31.0.dev0",
"use_cache": true,
"vocab_size": 32000
}

[mrwyattii]

Thank you for this contribution, but we have another PR to fix this: #4022

Closing this in favor of the previously submitted fix.