Need an efficient way to get CUDADeviceContext

[typhoonzero]

In an operator kernel, we usually got a framework::ExecutionContext object, and if we are implement GPU kernel, we often need to get CUDADeviceContext out from ExecutionContext. Currently we have to do sevaral cast to get it, need to simplify this.

Current sample:

auto* device_context =
        const_cast<platform::DeviceContext*>(context.device_context_);
auto cuda_ctx = 
         reinterpret_cast<platform::CUDADeviceContext*>(device_context);