Some misc virtualenv + library support

(dependencies of bcrypt)
Increase test timeouts since this takes around 90s to install
in debug mode.

Base off of sys.prefix instead of sys.executable, since in
virtualenvs the executable is in an extra bin/ directory.

But not both cmp and key at the same time.
We might want to just switch to CPython's sort implementation for that.

Remove a workaround in req_install.py

Rewrite calls into BoxedMethodDescriptors

This involves
- special-casing their __get__
- adding some ability to rewrite calls to functions that take *args

initial compile() support

Hopefully this helps detect more forms of file corruption,
such as 0-byte cache files.

Modules have a __file__ attribute but that's only used for
the module repr.  The filename that's used in tracebacks
is stored in the code object.

We don't currently destruct CLFunctions, but this makes it
clear that each SourceInfo belongs to a single CLFunction.

compile() is starting to work

Before, if we saw that they were going to immediately be executed in
the current globals scope (ie no "in" clause), we would execute them
without custom globals.

In order to split the compilation from the execution, just always
compile them with custom globals support.  This involved some refactoring
of the assumptions about when the globals object gets passed or not.

Using the makefile-to-cmake shim

Thanks to @dagar for building it for us!

Fix analysis issue that virtualenv was running into

The issue was that the types analysis was osr-aware, where
we would only type-analyze the sections of the function
accessible from the osr entry point.  The phi and definedness
analyses were not osr-aware, so they would think that phis
in certain places where the type analysis knew that they
were undefined.

So this change makes definedness and phi analysis osr-aware,
where they only analyze the appropriate section of the function.
I think this means that we will do these analyses more, since we
have to rerun them for each entry point, so hopefully analysis time
doesn't increase too much.

Metaserver misc

- fix list sorting gc bug (we did not hook std::get_temporary_buffer)
- eval(" 2")
- better SystemExit handling
- file.writelines
- list assign non-sequential slice

Speed-up analysis and irgen

this descreases the number of symbols / stackmap constants size slightly

because the time measurement takes more time than retrieving the cache entry

Take another pass over the CPython test suite

Jit cache misses

Speedup analysis by using llvms more cache friendly map/set algorithms

Iterate over the symbol table in sorted order.

I thought it would have been a problem with the -n optimization flag which we
would want to hunt down, but I think it might just be an issue with the parser.
Don't have a good idea why though.