[Prelude][Relay] Add get tensor data utilities

[wweic]

This implements issue #4291.

I also applied the optimization in tensorflow converter that it keeps track of all the ins and outs of every tensor array, so we can determine if a tensor array is homogeneous or not. If true, we will directly extract the tensor out of the tensor_t object, and it will eliminate the trouble of reading from tensor_t.

We should do the same optimization in the future when we encounter operators that needs to use tensor_t objects.

cc @zhiics @icemelon9 @yongwww @kevinthesun @srkreddy1238 @jroesch @slyubomirsky

[wweic]

I'll close the PR per discussion with @kevinthesun . We might want to detect static rank in the graph and specialize tensor array ops for the specific static rank, it will give most performant implementation. Generic tensor array might not be efficient enough for highly dynamic model.