Rescale tensors and choose correct contraction order optimizer in tests

[GiggleLiu]

In this PR, I

• Fix a bug about tensor shape,
• Rescale tensor networks to avoid exploding parameters, I do not have any good idea to tune the parameters automatically so far.
• Tune contraction order optimizer in the test,

Test Report

For "relational" dataset, none of the optimizers in OMEinsum find any proper contraction order. (there are more than 60k tensors, the size is too big I think).

The rest instances all pass except linkage_15 fails and ObjectDetection_35, I still do not have any clue about why these two instances are so special.

[GiggleLiu]

Update

Now I use the following solution to the rescaling problem:

1. create a new array type RescaledArray

   struct RescaledArray{T, N, AT<:AbstractArray{T, N}} <: AbstractArray{T, N}
   logf::T
   value::AT
   end

It represents array exp(logf) * value

2. implement einsum and some other function interfaces,

   for CT in [:DynamicEinCode, :StaticEinCode]
   @eval function OMEinsum.einsum(code::$CT, @nospecialize(xs::NTuple{N,RescaledArray}), size_dict::Dict) where N
       res = einsum(code, getfield.(xs, :value), size_dict)
       return rescale_array(RescaledArray(sum(x->x.logf, xs), res))
   end
   end

3. Then the overflow problem should disappear if we use this new array type for tensor network contraction.