SPIKE: Have ts2mod use suffwdpass and sufrevpass rather than sufrev

[awf]

Currently, given knossos function

(def foo ...)

We implement (in ts2ks.*_template) the forward pass by calling foo, and the reverse by calling [sufrev foo]. We should change to [suffwdpass foo] and [sufrevpass foo].

A current blocking issue, which @dcrc2 may have an idea on: pybind11

TypeError: Unable to convert function return value to a Python type! The signature was
        (arg0: dynamic_ksc_cpp.Tensor_1_Float) -> Tuple[dynamic_ksc_cpp.Tensor_1_Float, Tuple[Tuple[], Tuple[], ks::tensor<1ul, ks::tuple<double, ks::tuple<double, ks::tuple<double, ks::tuple<bool, double, ks::tuple<double, ks::tuple<bool, ks::tuple<double, ks::tuple<double, double> > > > > > > > >]]

[dcrc2]

The issue here is the tensor-of-tuples in the return type. We currently only support a small number of tensor types, listed in src/ksc/python/utils.py:

    declare_tensor_1<double>(m, "Tensor_1_Float");
    declare_tensor_2<double>(m, "Tensor_2_Float");
    declare_tensor_2<int>(m, "Tensor_2_Integer");

Clearly we can't just keep adding to this list, especially when we have complex implementation-defined structures like the one here. But there's also a further problem in that our implementation of declare_tensor_N only works for scalar element types anyway.

It's not actually clear to me what we would want this binding to look like. Do we want to copy the result into python objects, or create a python wrapper for a C++ object, or have some combination of the two depending on the element types? Maybe this needs some discussion/investigation.

Once we've decided what should happen I suspect it's possible to get pybind11 to do what we want, perhaps by implementing a type_caster specialization for ks::tensor. Or maybe ts2mod could code-generate the bindings. Either way it sounds like a significant chunk of work so I'd want to pin down the desired interface first.