Support lowering `fb-opt` from `onnx` to `torch` dialect

[rsuderman]

The following operations fail to lower with examples. These operations are supposed to be finished and will need to be corrected

• [x] onnx.Cast

%208 = torch.operator "onnx.Cast"(%207) {torch.onnx.to = 1 : si64} : (!torch.vtensor<[?,?,?,?],si64>) -> !torch.vtensor<[?,?,?,?],f32>

• [x] onnx.Div

    %227 = torch.operator "onnx.Div"(%222, %226) : (!torch.vtensor<[1,64,768],f32>, !torch.vtensor<[1,64,1],f32>) -> !torch.vtensor<[1,64,768],f32>

• [x] onnx.Equal

%205 = torch.operator "onnx.Equal"(%204, %175) : (!torch.vtensor<[4],si64>, !torch.vtensor<[4],si64>) -> !torch.vtensor<[4],i1>

• [x] #2702

%221 = torch.operator "onnx.ReduceMean"(%220) {torch.onnx.axes = [-1 : si64]} : (!torch.vtensor<[1,64,768],f32>) -> !torch.vtensor<[1,64,1],f32>

The following operations have no implementations:

• [x] onnx.Concat (this is done in this commit)
• [x] #2711
• [ ] onnx.Expand https://github.com/nod-ai/SHARK-Turbine/issues/302 #2714
• [x] onnx.Gather #242
• [x] onnx.Max merged
• [x] onnx.Mul merged
• [x] onnx.Pow merged
• [x] onnx.Reshape merged
• [x] #2696
• [x] onnx.Split merged

(after creating an issue to claim the above ops, please also register that you're working on it here in the big ONNX tracking issue)

[renxida]

Made a gist with some test cases for the ones with examples. Looks like Div already works as of today. I'll be tackling onnx.Equal.

[renxida]

Equal from this commit has an unecessarily high opset version (19) when i'm pretty sure we can support all the way down to 1

[stellaraccident]

I noticed a couple of others like that too. Would be good to do a scrub: https://github.com/llvm/torch-mlir/issues/2700