Features/clean prediction

[teubert]

This is an important PR, it is a breaking change and it includes the refactoring of the format of the results of prediction. The following changes are made:

1. Predictors are now expected to predict until every event of interest is met, or horizon is met, whichever is close. Can be specified by user, but default is all predicted events. Previously it stopped when first event is occurred
2. EOL is returned as key:value pair of {event_name: UncertainData (Unweightedsamples for MC)}.
3. Times is now returned as a 1d array - if an event has not occurred for a time, it is replaced with none (improving this behavior saved for Fix None behavior for end of prediction #79
4. So times align, final state at event occurrence is no longer part of inputs, outputs, states, etc. predictions. They will be added to the EOL prediction object in Save last state in EOL #80
5. Updated the examples that use metrics to fix break from this change

This PR includes the changes in MC to support this and a new example that predicts multiple events.

For clarity: Here are the changes the change:

Change 1: EOL format

Previous	New
predict from MC returns eol as array of doubles when first event occurs. Event is not labeled and may not be the same event (i.e., for sample 1 event 1 may occur before event 2, while it could be the opposite for sample 2	predict returns UncertainData. For MC that's UnweightedSamples where each event occurring is indicated by key: value pairs
Only 1 event	Any number of events
predict() -> [double]	predict() -> UnweightedSamples({event_name: time, ...})

Change 2: Times

Previous	New
times is in 2d array of format [sample_id][time_index] to support the non-aligned sample time steps	times is 1D array of times [time_index]. Biggest source of non-aligned sample times (state at event occurrence) removed and will be added into EOL as property. If a sample doesn't exist for a time, it is indicated by None

Note

Note 1: This requires changes in the unreleased version of prog_models, so released version tests will fail