WIP: Allow generation of multiple kernels simultaneously

This fixes #40 (closed) .

It allows the following syntaxes:

Generators have a new keyword argument context_tags=("foo",). The value of the tag foo is mangled into the cache key.

The value is taken from either the global context:

with global_context(foo="bar"):
    ...

or from a keyword argument to the generator itself, like:

@generator(foo="bar")
def bla()
    ....