Add support for Grouped Query Attention on Llama Model

[felladrin]

This change allows using models like ahxt/llama2_xs_460M_experimental.

• Resolves [Feature request] Support for ahxt/llama2_xs_460M_experimental model #388
• See @xenova's comment which defines the problem and points to the solution.

How to test

import { AutoModelForCausalLM, AutoTokenizer } from "http://localhost:8080/dist/transformers.js";
const model_path = "Felladrin/onnx-int8-llama2_xs_460M_experimental";
const model = await AutoModelForCausalLM.from_pretrained(model_path);
const tokenizer = await AutoTokenizer.from_pretrained(model_path);
const prompt = "Q: What is the largest bird?\nA:";
const input_ids = tokenizer(prompt).input_ids;
const tokens = await model.generate(input_ids, { max_length: 20 });
console.log(tokenizer.decode(tokens[0], { skip_special_tokens: true }));

Confirm it also works with pipeline:

import { pipeline } from "http://localhost:8080/dist/transformers.js";
const generator = await pipeline( "text-generation", "Felladrin/onnx-int8-llama2_xs_460M_experimental" );
const [output] = await generator("Once upon a time,", { max_length: 20 });
console.log(output.generated_text);

[xenova]

Thanks so much @felladrin! 🤗