implement missing operators

[johnynek]

See: https://github.com/onnx/models

It would be nice to have full coverage. From a few months ago, here is what is missing:

agate show zoo/*.onnx

TOTAL: (51 distinct ops, 3677 total ops, 24 unsupported(!) distinct ops (47.1%), 748 unsupported total (20.3%))
:
  Conv! (617), Relu (515), Unsqueeze (480), BatchNormalization (401), Add (282),
    Mul (230), Constant (200), Concat (139), Slice (101), Reshape (91),
    LeakyRelu (80), Cast (67), MaxPool (67), Squeeze (39), Transpose (39),
    Sum! (30), Div (29), Dropout (25), Gather (25), Shape (24), Sub (23),
    Gemm (20), AveragePool! (18), MatMul! (16), Softmax (16), Sigmoid (13),
    LRN! (12), Ceil! (10), ConstantOfShape (10), Loop! (6), Tile! (6),
    LSTM! (5), Pad! (5), ReduceSum (4), Exp (4), CategoryMapper! (4),
    ReduceMax! (3), Identity (3), GlobalAveragePool! (2),
    NonMaxSuppression! (2), Resize! (2), ReduceMin! (2), Log (2), TopK! (1),
    ArgMax! (1), Abs! (1), Clip! (1), Compress! (1), ImageScaler! (1),
    Scan! (1), Hardmax! (1)

So, Conv is the most common operator we don't support, followed by Sum, AveragePool, MatMul, LRN, Ceil...

Some operators are much easier than others. Sum and Ceil would be good first issues because they are relatively easy. MatMul isn't totally trivial, but we can call the existing Gemm code which is more general (just use a broadcasted 0 constant for the constant part).