[GPU / OpenCL] Moving buffer creation from inside of a layer to Tensor, i.e., update Tensor to support multiple ComputingEngine

[EunjuYang]

As Jijoong mentioned here, the current GPU handles 'opencl::Buffer' directly in all computations (every BLAS, layer operation, etc.). This can lead to several issues such as:

1. Incoherent interfaces between different compute units:
   • The current CPU computations utilize Tensors instead of memory buffers directly.
   • It would be more suitable if GPUs also used Tensors as input and output for their operations.
2. Latency and overhead for each computation:
   • Creating a buffer at runtime adds unnecessary latency.

[taos-ci]

cibot: Thank you for posting issue #2812. The person in charge will reply soon.