[DO NOT MERGE] Use only CUDA devices with a supported architecture

[fwyzard]

For each available CUDA device, check if its architecture is supported running a simple kernel.
This allows to restrict using only the supported devices

• at configuration time, in the SwitchProducerCUDA, via the cudaIsEnabled test;
• at run time, in the CUDAService and its clients;
• when running tests, via exitSansCUDADevices.

Includes Matti's fix for the sizes of stream and event caches (#289).

[fwyzard]

Second attempt at #286, mostly to check if we can find a working set of CUDA build rules for this use case.

[fwyzard]

The present CUDA rules do not allow to have .cu files both in a shared library and in a plugin.

[makortel]

Just to note (in case it was not intentional) that now there are implementations of supportedCUDADevices() in both HeterogeneousCore/CUDAServices and HeterogeneousCore/CUDAUtilities.

[fwyzard]

Sorry about that, I'll clean it up.

[makortel]

The logic here

cmssw/HeterogeneousCore/CUDACore/interface/CUDAESProduct.h

Line 18 in 586c0c0

CUDAESProduct(): gpuDataPerDevice_(numberOfCUDADevices()) {

cmssw/HeterogeneousCore/CUDAServices/src/numberOfCUDADevices.cc

Lines 5 to 8 in 586c0c0

	int numberOfCUDADevices() {
	edm::Service<CUDAService> cs;
	return cs->enabled() ? cs->numberOfDevices() : 0;
	}

cmssw/HeterogeneousCore/CUDACore/interface/CUDAESProduct.h

Lines 32 to 35 in 586c0c0

	edm::Service<CUDAService> cs;
	auto device = cs->getCurrentDevice();
	auto& data = gpuDataPerDevice_[device];

needs to be updated as well.