Ability to run different pipelines to different dispatches

[newling]

This task is to have an end to end model which has 2 dispatches, one with a convolution and one with a matmul. This will test our ability to run different pipelines automatically on different dispatches. An example is as follows (please excuse my python)

A = np.empty([64,64])
B = np.empty([64,64])
C = np.matmul(A,B)
D = C.reshape([1,64,8,8])
E = np.empty([1,8,3,3])
F = torch.conv2D(D, E)

This model will correspond to 2 dispatches in IREE, the first with a matmul the second with a convolution. There will be an expand_shape inbetween them.

To get this to generate code for AIE, the first dispatch (with the matmul) will need to use the pack-peel pipeline (or the pad-pack pipeline). The second dispatch (with the conv2d) will need to use the conv-decompose pipeline.

Currently the only way we can set the pipeline to run on the untiled function is manually with a flag. The compiler needs to dynamically choose the pipeline to run, based on the operations in the function. This must be a solved problem for all other backends that IREE has, no need to invent anything new I suspect.

The nearest thing I can see to an end-to-end test like this is the "three-matmuls" test: see here and here.

The task is to have an end-to-end numerical test of the above simple model.