56 decide on and implement dag solving strategy

[pabloitu]

• Implementation to run an experiment concurrently, with tasks have associated dependencies, and are placed into a directed acyclic graph. This mode is only available for time dependent models, whose source code is encapsulated in a docker image. In practice, catalog and evaluations tasks are sequential, whereas forecasts creation is run in parallel. User can set run_mode in the configuration file as sequential (default) or parallel.
• Decoupled engine runner from profiling. Profiler prints (in debug) a timing summary grouped by task class. Added catch for keyboard interrupts, so all running processes can be killed before exiting (critical for parallel running and container deployment).
• Minor rework of Task.str, such that it prints its instance name by default.
• Docker environments now properly catch errors and logs. Added also a static method, called from TaskGraph, that effectively kills all containers through the python SDK.
• To handle parallelization, models require to have a unique input folder for each time window. Therefore, a volume is set between a folder from the host filesystem (other than model/input), to /app/model/input within the container. Contents are located dynamically before running the container, by the methods CatalogRepository.set_input_cats, Model.prepare_args and Model.prepare_extra_input.
• Important fix on time_config parsing, which sometimes caused that time-dependent experiments were not recognized as such. Also, catalog repository now parses existing dropped catalog, rather than filtering it for each task.
• Updated all of the existing unit tests, extended engine and environments tests. created engine-model and engine-experiment integration tests. Added remaining tutorial cases (h,i,j) for end-to-end tests, which can only be run locally for processing time and space limitations in the CI.
• Added run_mode and force_rerun parameters in experiment configurations. Updated all refactored classes from last commits