[WIP] Add support for TupleGetItem node in Operator fusion

[masahi]

A partial fix for #2890. It can enable fuse through TupleGetItem node.

Example: GRU

%14 = fn (%X: Tensor[(1, 10), float32], %W: Tensor[(30, 10), float32]) -> Tensor[(1, 10), float32] {
  %1 = fn (%p0: Tensor[(1, 10), float32], %p1: Tensor[(30, 10), float32], __dict__=meta[StrMap][0]) -> Tensor[(1, 30), float32] {
    %0 = nn.dense(%p0, %p1, units=None)
    %0
  }
  %2 = %1(%X, %W)
  %12 = fn (%p01: Tensor[(1, 30), float32], __dict__=meta[StrMap][1]) -> Tensor[(1, 10), float32] {
    %3 = split(%p01, indices_or_sections=int64(3), axis=1)
    %4 = %3.0
    %5 = sigmoid(%4)
    %6 = %3.1
    %7 = tanh(%6)
    %8 = %3.2
    %9 = exp(%8)
    %10 = multiply(%7, %9)
    %11 = add(%5, %10)
    %11
  }
  %13 = %12(%2)
  %13
}

[tqchen]

Please also add an execution example to confirm we can generate correct code for cpu/gpu. Please tag more reviewers

[masahi]

Looks like my modification is not quite right. Will spend more time on it.