CenterPoint Backbone preprocessing optimization

[angry-crab]

The current implementation of scatter has some limitation.

1. the GPU implementation hard coded iterator bindings which might not work for certain devices. For example, for OpenCL backend, if a GPU has only one dimension global work size.

       for j in T.thread_binding(0, 560, thread = "blockIdx.x"):
           for k in T.thread_binding(0, 560, thread = "blockIdx.y"):
               for i in T.thread_binding(0, 32, thread = "threadIdx.x"):

2. There is no room for optimization because of hard code. Normally, we need to create schedule from IRModule and define optimization strategies.

3. Need to create a optimization schedule and measure its performance.

[angry-crab]

1 can be solved by implementing scatter from a top level, ie TE or Relay.