My theory is that specifying the separate cache directories
interacts badly with travis-ci.  I think it will either delete
any cache directories that look like they are no longer cached
(ie when it finishes the debug build it will delete the release
ccache dir since it looks unused), or maybe it keeps them around
but only checks the most recent build for a cache hit (ie the
debug build might check and only find ccache_release).

So anyway, try always specifying both directories.  Another
option is to do the simpler thing and use a single directory for
both caches, but I want to only change one thing at a time.

Slightly speedup ASTInterpreter::initArguments

Call finalizers during garbage collection with ordering

Make sure CPython functions like PyObject_Del do not free any objects.
It's better to let the GC free all objects because only the GC can
guarantee that there are no more references to an object.

New testing helper function ensures that objects are garbage
collected and their finalizers tested in a more strict order. This
reduces the chances that an object isn't collected because a reference
remains in the stack, but it's hard to fully solve this problem as long
as we have conservative stack scanning.

- Call all finalizers properly, in a safe-to-do-so order.
- Make finalizers work together with weak reference callbacks - namely,
  both should be able to get called, and weakref callbacks should be
  called before finalizers.
- Old style classes are also supported, but we always do an attribute
  lookup for tp_del because there's no tp_del slot. Since it would be
  too slow to do an attribute lookup all the time in GC, we just assume
  all old-style classes have an ordered finalizer.

Make sure __del__ methods get assigned to tp_del in new style classes.

Go through an attribute lookup in old style classes. Also, change the
function calling structure to match CPython's a bit better.

For exceptions thrown in tp_del methods, print a warning and ignore
them, like CPython does.

Fix parsing caching

Reenable attr interning

Flipped conditional meant that we were never caching the parse results :(

If we ask for a string as non-interned, and then ask for an interned
version of the string, we will discard the non-interned version and
generated a new interned version.  Now, keep the old version alive
as well.

From the course of debugging this gc issue.  this commit
shouldn't have much effect normally.

update libpypa

Add integration test for installing NumPy

travis-ci fixes

You can't use exit() to exit in the same way as if you had
died with a signal; we tried to do some exit-code munging
but we did it one way and our tester script expected another
way.  we could resolve that, or use the approach here: try
to die with the same signal by sending it to ourselves.

Doesn't work right now, but we'll want to support NumPy at some point.

The debug builds are taking a very long time to compile, but
when I reran one it completed very quickly.  Let's see if
this works+helps.

Add docs and tips for new contributors

try reverting these

This reverts commit 071ca0d8, reversing
changes made to aba975d1.

This reverts commit c78fdcb9.

Fixes

Instead, save the file data and parse that instead.
I think that should help with cases where the cached
file was getting trampled.

More travis-ci investigations

A fix to the ast interpreter gc visitor

more small perf work

Cache floats based on their bits, not their value

Fix some issues found during a search for another bug