LLVM / Torch-Mlir Upgrade

[zezhang]

• Remove StableHLO dependency
• Apply patches to build torch-mlir without stablehlo and stablehlo lit tests
• Update APIs realted to isa<>, cast<>, and dyn_cast<>
• Add a new EliminateUnusedTorchOpsPass to remove unused torch ops in torch-to-tcp (see example below):

module {
  func.func @func_main(%arg0: !torch.vtensor<[?,3],f32>, %arg1: !torch.vtensor<[?,3],f32>) -> !torch.vtensor<[?,3],f32> {
    %int2 = torch.constant.int 2
    %int0 = torch.constant.int 0
    %0 = torch.symbolic_int "s35" {min_val = 0, max_val = 9223372036854775807} : !torch.int
    torch.bind_symbolic_shape %arg0, [%0], affine_map<()[s0] -> (s0, 3)> : !torch.vtensor<[?,3],f32>
    torch.bind_symbolic_shape %arg1, [%0], affine_map<()[s0] -> (s0, 3)> : !torch.vtensor<[?,3],f32>
    %1 = torch.aten.size.int %arg0, %int0 : !torch.vtensor<[?,3],f32>, !torch.int -> !torch.int
    %2 = torch.aten.size.int %arg1, %int0 : !torch.vtensor<[?,3],f32>, !torch.int -> !torch.int
    %3 = torch.aten.sub.Tensor %arg0, %arg1, %int2 : !torch.vtensor<[?,3],f32>, !torch.vtensor<[?,3],f32>, !torch.int -> !torch.vtensor<[?,3],f32>
    torch.bind_symbolic_shape %3, [%0], affine_map<()[s0] -> (s0, 3)> : !torch.vtensor<[?,3],f32>
    %4 = torch.aten.eq.int %1, %2 : !torch.int, !torch.int -> !torch.bool
    %5 = torch.aten.Int.bool %4 : !torch.bool -> !torch.int
    %6 = torch.aten.Bool.int %5 : !torch.int -> !torch.bool
    torch.runtime.assert %6, "Runtime assertion failed for expression Eq(s35, s58) on node 'eq_2'"
    return %3 : !torch.vtensor<[?,3],f32>
  }
}

where the torch.runtime.assert and related checking ops can be removed.