[AutoTVM] RPCRunner network overhead

[FrozenGene]

When we use AutoTVM to tune on remote devices, we will introduce data copy to remote devices and bring network overhead: https://github.com/apache/incubator-tvm/blob/9d2c984ab633b372cd31054a5abd84a2e16f6220/python/tvm/autotvm/measure/measure_methods.py#L484

To solve this, we could implement a tvm.nd.non_empty in the tvm runtime, then we can do the random fill on the target device without copying over the network. I have implemented one version of it and plan to port it to upstream.

Open this to track this issue.