Use a common key-value store for simulation data

[benbovy]

Currently each Variable object hold its own data. With #19 this woudn't be the case anymore, so we need to find an alternative solution.

A common, writeable key-value store for all simulation data seems a nice and clean approach.

It could be as simple as a dictionary (in-memory store), e.g.,

store = {
    ('process_foo', 'param1'): 2,
    ('process_foo', 'var1'): np.array([2, 3, 4, 5]),
    ('process_bar', 'baz'): 4e-5
}

But it could be easily replaced by other kind of key-value stores, e.g., on-disk or distributed stores for simulations dealing with very large amount of data.

A new store would be created and assigned to each process of a model when initializing a simulation.

All properties created in each process (see #19) would therefore be simple object-to-store mappers.

[benbovy]

Closed in #33