Idea: rework Tile to reduce memory allocation overhead in ONNX

[Max-K-cadence]

Just an idea on what can be improved.

Currently, Glow's Tile node currently duplicates data along 1 dimension (and corresponds to such Caffe2 operator). ONNX's tile has a vector of values that show how many times we should multiply along every dimension. It is implemented by building a chain of Tile nodes that multiply data over and over again along each dimension. Whenever we create another node, we should create its output tensor, and this is where we use redundant memory.

We could rework Glow Tile node to have a vector of constants (with multiplication coefficients along each axis) to meet ONNX specification. With that, Caffe2 Tile is just a particular case of this operation with multiplication vector (1, 1, X, 1, 1). We save memory on creation of intermediate input/output tensors in ONNX Tile by doing so.

As a side effect, it enables the following optimization:

Y = Tile(X, (a, b, c))
Z = Tile(Y, (x, y, z))

into

Z = Tile(X, (a * x, b * y, c * z))

which, I believe, is not done now.