Graph Slice node (resolves #483)

[mreppen]

A node producing a get_slice on its input.

Most code that is not boilerplate is for computing the output dimension. The calculation itself is pretty much the same as in _enumerate_slice_def (https://github.com/owlbarn/owl/blob/master/src/base/dense/owl_base_dense_ndarray_generic.ml#L99). The assert is a bit more aggressive (checks bounds) to avoid Assert_failure ("src/base/core/owl_base_slicing.ml", 90, 10). when running.

I am aware of special cases like that "later" axes do not need to be specified, and handle that. Are there others that I might have missed?

There is one potential extra feature: A bool flag to set whether trivial dimensions (those with length 1) should be remove, i.e., reshaped away. Is this too niche/specialized?

There is still a TODO for prettier (more detailed) printing.

Example:

# let nn = Neural.D.Graph.(input [| 5; 2 |] |> slice [[1; -1; 3]] |> get_network);;
val nn : Owl_neural.D.Graph.network =
  42878

[ Node input_0 ]:
    Input : in/out:[*,3,2]
    prev:[] next:[slice_1]

[ Node slice_1 ]:
    Slice : in:[*,3,2] out:[*,2,2]
    prev:[input_0] next:[]

# Neural.D.Graph.model nn (Arr.sequential [| 1; 5; 2 |]);;
- : Owl_algodiff_primal_ops.D.arr =

       C0 C1
R[0,0]  2  3
R[0,1]  8  9

# Arr.(sequential [|5;2|] |> get_slice [[1;-1;3]]);;
- : Arr.arr =
   C0 C1
R0  2  3
R1  8  9

[mreppen]

I added more detailed printing:

# ... slice [[1; -1; 3]; []; [0]] ...
...
[ Node slice_1 ]:
    Slice : in:[*,5,2,10] out:[*,2,2,1]
    Axes  : 0:[1; -1; 3] 1:[] 2:[0]
    prev:[input_0] next:[]

[jzstark]

Cool! Thanks for this PR! Looks good to me. I'm just wondering if we can put the slicing shape inference routine to the owl_utils_infer_shape.ml file.

[mreppen]

@jzstark I was not aware of that file, but that looks very appropriate. Latest commit moves it.