[JIT] Docstrings for custom C++ classes

[jamesr66a]

We should have a way to specify docstrings on a) custom registered classes and b) the methods thereof. I believe this will entail:

1. C++ APIs for populating the docstring. I can't find a pybind11 equivalent for installing a docstring on a class, but it could be an extra string argument to the torch::class_<> constructor. For method definitions I think it should similarly be an extra string argument to class_::def()
2. A way to represent docstrings in the TorchScript type system. For classes, it could preferably be a __doc__ attribute, or a special case method on the ClassType:

   pytorch/aten/src/ATen/core/jit_type.h

   Line 1950 in 40828b6

   struct CAFFE2_API ClassType : public NamedType {

   . For methods, I think we would need to add a special docstring method to jit::Function:

   pytorch/aten/src/ATen/core/function.h

   Line 27 in 40828b6

   struct TORCH_API Function {

3. Then, in the Python bindings, for classes we would need to bind in a __doc__ property to access the docstring as we defined in step (2):

   pytorch/torch/csrc/jit/python/python_custom_class.cpp

   Line 31 in 40828b6

   .def("__call__", &ScriptClass::__call__);

   . For functions similarly we would bind in the __doc__ property on ScriptFunction:

   pytorch/torch/csrc/jit/python/script_init.cpp

   Line 1111 in 40828b6

   py::class_<StrongFunctionPtr>(m, "ScriptFunction", py::dynamic_attr())

Note that for any of these stages, if we can generalize the support to broader TorchScript support (particularly the frontend concerns), that would probably be a good idea

cc @gmagogsfm

[suo]

Requested by @fmassa

[gmagogsfm]

@wanchaol

I think this would be useful for torch.distributed as well.

[suo]

how pybind11 API looks: https://pybind11.readthedocs.io/en/stable/advanced/misc.html#generating-documentation-using-sphinx