`PyArray_DescrConverter_` doesn't steal a reference to the argument,
and so the passed arguments shouldn't be `.release()`d.

`type_descr` is now applied only to the final signature so that it only
marks the argument types, but not nested types (e.g. for tuples) or
return types.

MSCV does not allow `&typeid(T)` in constexpr contexts, but the string
part of the type signature can still be constexpr. In order to avoid
`typeid` as long as possible, `descr` is modified to collect type
information as template parameters instead of constexpr `typeid`.
The actual `std::type_info` pointers are only collected in the end,
as a `constexpr` (gcc/clang) or regular (MSVC) function call.

Not only does it significantly reduce binary size on MSVC, gcc/clang
benefit a little bit as well, since they can skip some intermediate
`std::type_info*` arrays.

The current C++14 constexpr signatures don't require relaxed constexpr,
but only `auto` return type deduction. To get around this in C++11,
the type caster's `name()` static member functions are turned into
`static constexpr auto` variables.

When Travis changes their default Python 3.x, it breaks any hardcoded
version selection. Fix: make pyenv activate everything (2.7, 3.x) and
use whichever Python 3.x is on by default.

[skip appveyor]

* Expand documentation to include explicit example of py::module::import 
  where one would expect it.

* Describe how to use unbound and bound methods to class Python classes.

[skip ci]

E.g. trying to convert a `list` to a `std::vector<int>` without
including <pybind11/stl.h> will now raise an error with a note that
suggests checking the headers.

The note is only appended if `std::` is found in the function
signature. This should only be the case when a header is missing.
E.g. when stl.h is included, the signature would contain `List[int]`
instead of `std::vector<int>` while using stl_bind.h would produce
something like `MyVector`. Similarly for `std::map`/`Dict`, `complex`,
`std::function`/`Callable`, etc.

There's a possibility for false positives, but it's pretty low.

Fixes #1069.

Fixes #1075.

`PyNumber_Float()` and `PyNumber_Long()` return new references.

To avoid an ODR violation in the test suite while testing
both `stl.h` and `std_bind.h` with `std::vector<bool>`,
the `py::bind_vector<std::vector<bool>>` test is moved to
the secondary module (which does not include `stl.h`).

[skip appveyor]

This runs the most basic tests first and avoids waiting until the very
end for style checks.

[skip appveyor]

* Update Python 3 osx image to xcode8.3 to speed up brew install. 
  The Python 2 osx image remains xcode7.3.

* Have one osx config run in debug mode to improve coverage.

* Only run CMake build tests on two configs to speed up overall build.

  The CMake tests take ~30 seconds on each configuration, but we really 
  only need to them to run on two: one on Linux and one on macOS. This
  mirrors the recent change on AppVeyor.

* Merge the style/docs/pip tests with the barebones build.

* Merge 32-bit and CMake install configurations.

  This removes clang 3.9 from testing, but there are already 3 other clang 
  versions being tested on Travis and the new xcode8.3 image should be 
  close to clang 3.9.

[skip appveyor]

[skip ci]

Fixes #1061.

`T` and `const T &` are compatible types.

Closes #1048, closes #1052.

[skip ci]

[skip ci]

PR #880 changed the implementation of keep_alive to avoid weak
references when the nurse is pybind11-registered, but the documentation
didn't get updated to match.

[skip ci]

Fixes #1046.

[skip ci]

There are two separate additions:

1. `py::hash(obj)` is equivalent to the Python `hash(obj)`.
2. `.def(hash(py::self))` registers the hash function defined by
   `std::hash<T>` as the Python hash function.

The warning is shown at module initialization time (on import, not
when the functions are called). It's only visible when compiled in
debug mode.

This is analogous to `py::init()` vs `__init__` + placement-new.
`py::pickle()` reuses most of the implementation details of `py::init()`.