[ANSOR] Auto-scheduler tutorial for GPU and necessary refactor/fix

[merrymercy]

• Add a tutorial on auto-scheduling a subgraph for GPU (tutorials/auto_scheduler/tune_conv2d_layer_cuda.py)
• Refactor evolutionary search
• Fix MutateComputeLocation
  • In the old implementation we reuse InitChangeComputeLocation, but this is wrong. Because it always appends new steps and will make the transform history too long. In the new implementation, we only mutate the old steps, so the number of steps keeps the same.

[merrymercy]

cc @jcf94 @comaniac @junrushao1994 @binarybana

[jcf94]

This may still be confusing. I'm thinking it may be better to just delete the resume part of the CPU tutorials to keep it as simple as possible.

We can add a comment saying that LocalRPCMeasureContext is recommended to be used in all hardware targets in the GPU tutorial.

[merrymercy]

People are likely to get into this problem in their applications, so it is worth to let them aware of this problem in the tutorial.

[merrymercy]

Actually, I removed LocalRPCMeasureContext to make the major part of the tutorial simpler.
Because resuming a search is a very advanced usage, most people don't need this part. So we can make it complicated. But all sections above should be as simple as possible.

[comaniac]

Intuitively, if there's no obvious performance impact, we should use RPC runner on both CPU and GPU, so it'd better to mention why we didn't use it in this tutorial.

[comaniac]

I missed the comment from @jcf94 when reviewing this part. I agree that this may be confusing, but I prefer to keep the resume part; otherwise users will still encounter this issue when they are trying to resume the search. For example, one may write a script to do the following:

sch, args = do_search(task, log_file)
perf = evaluate_result(sch, args)
while perf < goal:
  sch, args = resume_search(task, log_file)
  perf = evaluate_result(sch, args)

[merrymercy]

Comments are addressed. @comaniac @jcf94 @tqchen

[comaniac]

Thanks @merrymercy @jcf94