New context switching code for generators

Smaller performance improvements for fasta

remove the larger buckets, and hoist some math out of loops.

Python exposes the finding part through the 'imp' module.

It uses the buffer protocol, so make str support that better.

For some reason the larger bucket sizes are causing a large perf hit
in spectral_norm.  It's unclear exactly why this is happening, but
theories are legion.  More investigation is warranted, but this gets us
back from the perf regression.

Also hoist the atom_idx calculation out of a couple of loops that were
iterating over object indices.

Only gets hit when there are >=3 !is_defined names also set (other
fake names might also count towards this).

Don't emit duplicate attr guards

pyston (calibration)                      :    0.8s stock2: 0.8 (+2.5%)
pyston interp2.py                         :    5.9s stock2: 6.2 (-4.5%)
pyston raytrace.py                        :    6.9s stock2: 7.0 (-1.6%)
pyston nbody.py                           :    9.8s stock2: 9.6 (+1.9%)
pyston fannkuch.py                        :    7.0s stock2: 6.9 (+2.6%)
pyston chaos.py                           :   20.6s stock2: 21.6 (-4.6%)
pyston spectral_norm.py                   :   27.9s stock2: 34.2 (-18.6%)
pyston fasta.py                           :   17.1s stock2: 17.8 (-4.5%)
pyston pidigits.py                        :    4.4s stock2: 4.5 (-1.0%)
pyston richards.py                        :   10.4s stock2: 10.2 (+2.2%)
pyston deltablue.py                       :    2.2s stock2: 2.2 (-1.9%)
pyston (geomean-0b9f)                     :    8.8s stock2: 9.1 (-3.2%)

reduces the generator yield overhead

This is a huge speed improvement for generators,
fasta.py takes 8secs now instead of 18secs

reduces strJoin runtime from 0.8sec to 0.5sec when executing fasta.py

We should do a more comprehensive investigation.  Removing t2 caused
regressions on a number of benchmarks since we lost chances to do
speculations, but making t3 easier to get to caused regressions
due to the cost of our LLVM optimization set (which is pretty hefty
since it's supposed to be hard to activate).

Fully switch to the new deopt system, and clean up a lot of stuff.

A "FunctionSpecialization" object really makes no sense in the context of
an OSR compile, since the FunctionSpecialization talks about the types
of the input arguments, which no longer matter for OSR compiles.
Now, their type information comes (almost) entirely from the OSREntryDescriptor,
so in most places assert that we get exactly one or the other.

We only needed that for supporting the old deopt system

Long live new-deopt!

Before we would do type analysis starting from the function entry
(using the specialization of the previous function).  This makes things
pretty complicated because we can infer different types than we are OSRing
with!  Ex if the type analysis determines that we should speculate in an
earlier BB, the types we have now might not reflect that speculation.

So instead, start the type analysis starting from the BB that the OSR starts at.
Should also have the side-benefit of requiring less type analysis work.

But this should let us get rid of the OSR-entry guarding, and the rest of
the old deopt system!

add a simple_destructor for file objects

Add third GC arena

Previously it was:
tier 0: ast interpreter
tier 1: llvm, no speculations, no llvm opts
tier 2: llvm, w/ speculations, no llvm opts
tier 3: llvm, w/ speculations, w/ llvm opts

tier 2 seemed pretty useless, and very little would stay in it.  Also,
OSR would always skip from tier 1 to tier 3.

Separately, add configurable OSR/reopt thresholds.  This is mostly for the
sake of tests, where we can set lower limits and force OSR/reopts to happen.

Generator tracebacks

add reversed() builtin function, as well as compatible __reversed__ methods on list/xrange.

Fix: GC must visit the 'iter' field of every BoxedIterWrapper