FIX: Add safe guards for static cache + llama on transformers latest

[younesbelkada]

Fixes: huggingface/transformers#28470

This PR makes autoawq + transformers compatible with recent changes on Llama architecture. In transformers llama the causal mask is pre-allocated with the size bsz, 1, max_seq_len, max_seq_len, thus needing to slice the attention mask here in the fused attention module.

Also fixes an issue where running this script fails:

Script to repro:

from transformers import AutoModelForCausalLM, AwqConfig, AutoTokenizer

awq_config = AwqConfig(do_fuse=True, fuse_max_seq_len=512)
model = AutoModelForCausalLM.from_pretrained(
    "casperhansen/tinyllama-1b-awq",
    quantization_config=awq_config,
).to("cuda")

tokenizer = AutoTokenizer.from_pretrained("casperhansen/tinyllama-1b-awq")
input_ids = tokenizer("Hello, my dog is cute", return_tensors="pt").input_ids.to("cuda")

model.forward(input_ids, use_cache=False)
model.generate(input_ids, max_new_tokens=100)

We defer to not using the caching logic if use_cache=False

cc @casper-hansen