use a binary search to find/insert CompiledFunction*'s into the registry

colocate string and tuple data

Tuples now contain a pointer to their storage (Box**) + size, and strings contain
an llvm::StringRef which contains the pointer+size.

use the same hash function for both unicode and string objects

this keeps us from having to utf8 decode unicode objects (into BoxedStrings)
whenever we want to hash them.

free generator stacks both in the destructor and when a generator exits.

we're already unmapping the redzone as part of the munmap() call.  also, free the generator stack from the sending side, not from within the generator.

a few changes which help out GC performance

when looking up a child hidden class we look up the attr in the child_index map,
then get the HiddenClass* from the children vector (at that index.)

this speeds up GC substantially by letting us visit the range from the vector as
opposed to iterating through the stringmap.

add a map-like container that keeps its values contiguous so we can use the efficient gc_visitor->visitRange.

also, don't register MAX_STACK_SIZE bytes with the GC, register INITIAL_STACK_SIZE.

Somewhat tricky since it assumes __dict__ is always a dict object;
I think I caught most of the places though.

implemented kmod's suggestions

through syntax error when you to `del` a var stored in a closure

add option to tester script to skip specified tests

Conflicts:
	src/codegen/ast_interpreter.cpp
	src/runtime/types.cpp

Update INSTALLING.md

add "sudo" before apt-get
It seems a typo?

add StackRoot<> and #defines for DECREF

add a variadic template ctor that delegates to T's constructor (making things potentially a bit more compact for constructing rooted boxes)

Misc fixes for virtualenv / pip