[Performance] Change the shape of kv_cache to avoid view of k_cache and v_cache. (#204)

This PR changes the shape of kv cache to avoid the view of k_cache and
v_cache.
What's more, cache the metadata of k_cache and v_cache to avoid
duplicative slice operations to improve performance.

Signed-off-by: hw_whx <wanghexiang7@huawei.com>

vllm_ascend/worker/worker.py

@@ -266,6 +266,12 @@ class NPUWorker(LocalOrDistributedWorkerBase):
                         self.parallel_config, self.device_config)
             for _ in range(self.parallel_config.pipeline_parallel_size)
         ]
+        import torch_npu
+        for ve in range(self.parallel_config.pipeline_parallel_size):
+            num_layers = len(self.cache_engine[ve].gpu_cache)
+            for i in range(num_layers):
+                torch_npu.npu_format_cast(self.cache_engine[ve].gpu_cache[i],
+                                          2)
         self.gpu_cache = [
             self.cache_engine[ve].gpu_cache
             for ve in range(self.parallel_config.pipeline_parallel_size)