Add support for global models

[edtechre]

Currently models can only be trained per symbol. Several users have suggested making it possible to train a model on data for multiple symbols.

[edtechre]

Not (easily) feasible given that dates may not align for different symbols. This would be confusing to users and there is no good way to unify semantics here like with training models per-symbol. And there is no canonical way to train on multiple symbols given overlapping/non-overlapping dates. i.e. do we simply dump a dataframe of all dates, bar data, and indicator values for symbols, even if the dataset does not have overlapping dates or is unbalanced?

It would be better if users train their own model, store it in pybroker.param and then use set_before_exec or set_after_exec handlers to pass input to the model for multiple symbols.