Render NaN default argument as numpy.nan

[sizmailov]

This PR changes how nan values are rendered in docstring.

this PR	master
foo(x: float = numpy.nan) -> None	foo(x: float = nan) -> None

Implemented via float overloads for py::arg_v. Might be considered as an overkill for such a small change. On the other hand nan values are quite common sentinel values so it might be worth it to generate more toolable docstrings.

compare.cpp

#include"pybind11/pybind11.h"
#include"pybind11/embed.h"namespacepy= pybind11;
PYBIND11_EMBEDDED_MODULE(example, m) {
m.def("foo", [](float x) {}, py::arg("x")=std::numeric_limits<float>::quiet_NaN()); }
intmain() {
py::scoped_interpreter guard{};
py::exec(R"( from example import *  print(foo.__doc__, end="") )");
}

[sizmailov]

__cplusplus >= 201103L doesn't mean the library is c++-11 complaint in gcc-4.8 (std::isnan is missing).
To avoid mess with other compilers versions I've enabled it starting from c++14.

[sizmailov]

On second thought I think I chose wrong place inject the change.

The underlying bigger problem is __repr__ function is not intended to produce valid python expressions, therefore we get malformed function signatures quite often. Probably next suggestion is better candidates to format value for signatures.

pybind11/include/pybind11/pybind11.h

Lines 210 to 213 in eeb1044

	if (a.descr)
	a.descr = strdup(a.descr);
	elseif (a.value)
	a.descr = strdup(a.value.attr("__repr__")().cast<std::string>().c_str());

For example new extension point can be introduced with something like

tempate<typename T>
structvalue_formatter {
value_formatter(const T& value);
constchar* operator() const;
}// and later in pybind11.hif (a.descr)
a.descr = strdup(a.descr);
elseif (a.value)
a.descr = strdup(value_formatter(cast(a.value))());

[wjakob]

I'm curious: what's wrong with just nan? I don't think numpy.nan is a good alternative, pybind11 is used in all sorts of places that don't even have numpy installed.

[sizmailov]

The problem is that generated signatures are not valid in sense of interpretation as stubs signatures. nan refers to a name which is not defined, so docgen/static analysis tools gets confused. The numpy.nan can be replaced with float('nan'), but I think it's a bit more readable.

The situation is worse when default value is not even valid python expression, e.g. x: Foo = <Foo instance 0xffff>. I think it needs generic solution so I close this PR. In shed of light #2244 closing I think you will argue against any PRs in this direction too.

Documentation matters therefore I think it's worth to produce toolable docstrings.