Improve work-to-device scheduling

[makortel]

This issue is about operating multiple GPUs.

In HeterogeneousEDProducer the (work in) EDM streams is assigned to devices in a "static" way, i.e. each producer on each EDM stream will always use the same device.

In #100 the logic stays (for backwards compatibility), but after all producers have been migrated from HeterogeneousEDProducer to the model of #100, we could try more clever load balancing between the devices.

[makortel]

One question is that whether we schedule each event fully to one device, or allow producers of a single event to use (or put their output product on the memory of) different devices.

Former is certainly simpler to start with.

Latter needs a model for reading a data product from the memory of another device. This can be trivially achieved with unified memory (#85). In #100 (comment) @fwyzard commented also

Often, we could let the currentDevice_ read the memory of the dataDevice over the PCI-e or NVLINK bus.

[fwyzard]

Just for the record, I'm trying to think what benefits we could find in running part of the event on different devices (of the same type).

A simple use case (that we are very far from) is if a single event is enough to saturate and exceed a single device.

An other is, if we end up being memory limited, we could keep some conditions (e.g. pixel calibrations) on one gpu, and others (e.g. calorimeters calibrations) on a second one, and run the relevant algorithms on the corresponding device.

Both seem far off.