Revisiting library initialization

[magnatelee]

The current procedure for library initialization requires that the library registration callback be exposed and bound to Python via CFFI. This shouldn't be a hard requirement, as long as the library makes sure that its registration callback is invoked somehow. Plus, while working on #675, @trxcllnt and I uncovered that the dlopen during library registration can cause trouble when mixed with Cython extension libraries that are misconfigured, such as shared libs of the same legate library from different locations being loaded simultaneously.