Generators: Support explicit StopIteration raising.

Before this patch we couldn't run 'list(G())' if G contained a explicit raise StopIteration statement.
This also adds support for custom StopIteration subclasses.

unwinder

The signal-blocking is used to prevent libunwind from being entered
recursively, if a signal handler invokes libunwind. We don't call
libunwind from signal handlers, so this is not a problem for us.

The conservative checks just make crashes behave more nicely. They have
a small but measurable performance impact. We enable them in debug mode
and disable them for release mode.

We also modify .gitmodules so that our patches to libunwind/ don't make
git think the tree is dirty.

http://blog.jgc.org/2015/04/the-one-line-you-should-add-to-every.html

Perf investigations

The issue is that sometimes we want to test out using builds
that wouldn't be able to self-host the libraries that the build
system thinks we need.  So, add a flag to just not build those.

I think this only matters if we knew the type at JIT time.

also, improve the annotator

Add alloc stats to the base Box::operator new(), which is where non-builtin
types get allocated.

Also add stats for non-Python allocations (typically, conservative memory);
turns out that we are allocating a lot of conservative memory (both in terms
of allocations but also in terms of allocation size) for dicts.

Hopefully this makes it more clear which stats are subsets of others,
and which ones should add up to other ones.

ex it was easy to think that "rewriter_attempts" succeeded as long
as "rewriter_aborts" didn't get incremented, but it could also not
have gotten started, etc.

Here's an example of the new stats:
ic_attempts: 1724625
ic_attempts_nopatch: 1306007
ic_attempts_skipped: 404969
ic_attempts_skipped_megamorphic: 85558
ic_attempts_successful: 13649
ic_rewrites: 13649
ic_rewrites_aborted: 3248
ic_rewrites_aborted_assemblyfail: 65
ic_rewrites_committed: 10401

implement float.hex and float.fromhex

remove the llvm::StringRef from BoxedString

support arg unpacking

tracebacks, struct.pack

The struct module just defers to PyLong_As* methods, so the
actual change here is improving our implementation of
_PyLong_AsByteArray, to add
- overflow and underflow checking
- twos-complement conversion (mpz_export ignores the sign)
- endianness

Now that we have frame support including f_back, we can
use the cpython implementations of these functions.

We don't have the same traceback objects (don't have
tb_frame or tb_next), so there are still some modifications
to the traceback library.

Copy all HAVE_FOO defines from cpython

Previously we were only copying them when we noticed they were needed,
but we've been spending a fair bit of time hunting down issues that
are caused by a missing HAVE_ define.

So instead, bulk-copy the ones that the CPython configure script determined.
I bet we'll start running into some that were available on my 14.04
machine but won't be on other configurations; I'm hoping that an extra
define is easier to debug than a missing one (build error vs run error),
making this the better approach.

Add full_check and fix a couple misc things