Support more general tensor comparisons

[gbaydin]

Currently we have the behavior

let t = dsharp.tensor([1.;2.;3.])
t.lt(dsharp.tensor(3.))

System.ArgumentException: The arrays have different lengths.
array1.Length = 3, array2.Length = 1 (Parameter 'array1')

We can generalize these comparisons to support cases like the following in PyTorch

t = torch.tensor([1,2,3])
t.lt(torch.tensor(3))

tensor([ True,  True, False])

[gbaydin]

Another behavior that can be updated in this is

dsharp.tensor([1.;2.;3.]) < 3.

This expression was expected to have type
    'Tensor'    
but here has type
    'float'    

dsharp.tensor([1.;2.;3.]) < dsharp.tensor(3.)

System.Exception: Cannot compare non-scalar Tensors

instead of which we can again support more general comparisons like PyTorch

torch.tensor([1,2,3]) < 3

tensor([ True,  True, False])

I think this can be accomplished by overloading the relevant operators (op_LessThan etc.).

[dsyme]

instead of which we can again support more general comparisons like PyTorch

Annoyingly the < and other comparison operators are tied to a return type of bool in F#. It's something we should address. Instead one has to make a <. or other operator specifically for tensors I think. Though that's not such a bad thing.

[gbaydin]

Annoyingly the < and other comparison operators are tied to a return type of bool in F#.

Oh good point, I remember now that this was the main reason behind this choice.

I think then just supporting the lt, le etc. with multi-dimensional bool tensor results is good enough.

[kevmal]

I think there's no harm in adding the .< style ops to the type. It would actually feel natural to have A = B be bool and A .= B to be a bool Tensor. In general it greatly improves readability over lt, lte style methods.