Add TemporalDataStore

[aurorarossi]

This PR implements TemporalDataStore.

[CarloLucibello]

A few comments:

• Any particular reason for having time as the last dimension instead of nodes? Node dimension in the GNN context is similar to the batch dimension. In the Flux ecosystem sequences (e.g. in NLP) have shape [embed_dim, seq_len, batch_dim], so here we probably want to have [embed_dim, t, num_nodes] and [embed_dim, t, num_edges].
• Do we need a TemporalDataStore structure at all? Can't we just use DataStore, since it can store arbitrary dimension arrays already? We could extend it to optionally check the size of the second to last dimension as well.
• Is it enough to have a data store that assumes that features exist at all times, or do we need something more flexible?