Agnostic Engine configured by specific handlers

[sdesrozis]

🚀 Feature

As discussed https://github.com/pytorch/ignite/pull/1919#issuecomment-821959142, some specific features are hardcoded in the Engine (reset of batch/output for instance). Usually it is internal concerns and therefore well encapsulated in the Engine.

However, the Engine is a use-agnostic concept that aims to iterate collections (typically datasets) by applying a function (i.e. the training_step) and providing an Event handler for extensibility. In order to offer even more flexibility, the idea here is to provide an internal configuration of the Engine using specific internal-use handlers.

The general aim is to have the most generic engine possible and to make maximum use of the Event mechanics associated with the Engine.