[BugFix]add int8 cache dtype when using attention quantization

[Angazenn]

1. Ascend attention requires int8 kvcache dtype. It is used in initialization of CacheConfig:

if cache_config.cache_dtype == "auto":
    self.dtype = model_config.dtype
else:
    self.dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_config.cache_dtype]

STR_DTYPE_TO_TORCH_DTYPE is defined in vllm.utils:

STR_DTYPE_TO_TORCH_DTYPE = {
    "half": torch.half,
    "bfloat16": torch.bfloat16,
    "float": torch.float,
    "fp8": torch.uint8,
    "fp8_e4m3": torch.uint8,
    "fp8_e5m2": torch.uint8,
}

Hence we need to update both cache_dtype and STR_DTYPE_TO_TORCH_DTYPE.

2. In vLLM, attention quant methods are initialized by passing self of layer. When initializing ascend attention quant method, num_heads, head_size, num_kv_heads are required. However, in the codes of Attention.init from vLLM, these member variables are set after initialization of quant method. Hence we have to move these codes ahead of quant method.

# move ahead of quant method.
self.num_heads = num_heads
self.head_size = head_size 
self.num_kv_heads = num_kv_heads

quant_method = quant_config.get_quant_method(
    self, prefix=prefix) if quant_config else None

[whx-sjtu]

Does vllm originally has entrypoints to resgister custom quant types into STR_DTYPE_TO_TORCH_DTYPE? If not, maybe we need to make this an issue to vllm? @wangxiyuan please check this.

[wangxiyuan]

No, we can create a pr to vllm later.

Currenlty. for this kind of change(Monkey patch), please move it to the patch module.

[MengqingCao]

CI failed due to the test path update in vLLM, will fix in #124