* If the last refcount on a python lambda/temp function is released when a
wrapping std::function is destroyed in a C++ invoke, we can't destroy it
safely. Instead we queue up that decref and process it the next chance we're
able (which is either when the current execution finishes for a python shell
execution, or on the next function call which handles extensions).
* When calling the register() function there is no frame or globals, so we need
to set an internal handle external to that.
* This means functions that get wrapped know that there's a global handler for
exceptions, which just prints to the log. Otherwise they think they're running
synchronously and write to an invalid exception handling object.
* This will rarely be relevant, but it ensures if the function is decref'd and
the lambda is still alive, that we keep the python object alive until we are
done with it. The primary case for this is persistent callbacks where the
module is then reloaded.
* Previously we would convert from python to C++ arrays immediately by
copying, and vice-versa convert TO python immediately by creating a
new python list by copying.
* This however behaves rather poorly in common situations, e.g.:
> foo.bar.append(5)
Would not append 5 to foo.bar, but copy foo.bar to a temporary, append
5 to it, then destroy it leaving foo.bar untouched.
* Instead we leave the C++ array type as a pointer for as long as we can
and instead implement the python sequence API as extensions/slots that
work in-place on the original array.
kb1000