Try to autoinstall manylinux1 wheels when possible

[jamesls]

In order to create a lambda deployment package, chalice currently will:

• Create a virtualenv
• pip install your requirements files into the virtualenv
• copy over the chalice app, chalicelib, and vendor directory into site-packages
• zip up the site-packages directory

This process works as long as all the packages are pure python packages. If you specify something like numpy, we'll install the packages based on whatever platform you're running on, which is problematic. For example, on a Mac, we'll install the numpy binaries for mac, zip them up and send them over to lambda, and import numpy will fail when the lambda function is invoked.

We've received a number of issues about this:

• Importing the multiarray numpy extension module failed #239
• Error: Unable to import your app.py file: numpy.dtype has the wrong size #105
• {"errorMessage": "Unable to import module 'app'"} #106

What I'd propose is a best-effort attempt at using manylinux1 wheels:

• For each package, if there's a manylinux1 wheel available, then use it
• Otherwise if there's a suitable non-platform specific wheel, then use it
• Otherwise if it's a pure python package, just fall back to the normal sdist install
• Otherwise if it requires C extensions and we don't have a manylinux1 available, fail (or maybe just warn).

I'm not sure how feasible it is to implement the above logic, but that would be my ideal behavior.