Modules being imported are now detected at compile time. So it's
possible to determine whether "from x import y" is importing module y or
a member y of module x. This change uses this capability to choose
whether to import a module or a member.

This fixes https://github.com/google/grumpy/issues/99

Previously, grumpc assumed that all imports were valid Python modules.
So there was no way to distinguish between, for example, importing
a module within a package, or a package member of the same name.

This change makes it so that during code generation, Python source files
corresponding to the modules being imported exist on the GOPATH. If no
such Python source file exists, then an ImportError is raised.

Additionally, this change fixes an issue where imports within packages
or within modules within packages were not first checking for the named
import within the package directory. Now, package-relative imports work
as expected.

Import logic is currently embedded in the StatementVisitor class which
makes it difficult to reuse. This change factors that logic out into an
ImportVisitor class in util.py. The newly exposed logic can now be used
by other tools.

Python code compiled by grumpc will be staged in GOPATH/__python__ so
that modules can be detected at compile time.

Previously add_import(), add_native_import() and intern() were Block
methods that delegated to the root ModuleBlock internally. Now that the
ModuleBlock is accessible from any Block via the root attr, move those
methods to ModuleBlock.

Per the docs, mktime expects a time struct in local time, not UTC:
https://docs.python.org/2/library/time.html#time.mktime

This mismatch was causing TestDate.test_fromtimestamp in
test_datetime.py to fail on some systems.

dictNE was checking v.isInstance(DictType) instead of checking w. This
causes a variety of undefined behavior including segfault, invalid mutex
state, etc.

Addresses https://github.com/google/grumpy/issues/56