fix the issue of hipblas matmul

[zhangnju]

when running the int8 benchmark of official bitsandbytes https://github.com/bitsandbytes-foundation/bitsandbytes/blob/main/benchmarking/int8/int8_benchmark.py , or the below sample codes for int8 quantization:

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

base_model_name = "/models/Llama-2-7b-hf"
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
quantization_config = BitsAndBytesConfig(load_in_8bit=True)
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
bnb_model_8bit = AutoModelForCausalLM.from_pretrained(
base_model_name,
device_map="auto",
quantization_config=quantization_config)

prompt = "What is a large language model?"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
generated_ids = bnb_model_8bit.generate(**inputs)
outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
print(outputs)

we met the below hipblaslt matmul issue:

this patch can fix the above issue by 1) removing rocblas in class context, since only hipblaslt is used in rocm bitsandbytes 2) adding the workspace pointer for hipblaslt matmul. this patch also includes some more log info to help debug.
after applying this patch, both the above sample codes and the int8 benchmark codes can run successfully.