Calculate paddle_peak_increase using paddle_allocated_mem_after_run：
paddle_peak_increase = paddle_allocated_mem_after_run - paddle_allocated_mem_before_run
This way paddle_peak_increase is closer to the actual situation.

The reserved variable returns the current memory size managed by the Allocator. The allocated variable returns the current memory size allocated to the Tensor. The difference we are concerned about here is the memory size allocated to the active Tensor before and after the profile. Therefore, theoretically, using paddle_allocated_mem_after_run - paddle_allocated_mem_before_run can meet the requirement. Using reserved may cause the calculated difference to be artificially high, resulting in a smaller memory size allocated to the kv-cache