Fix CachingHostAllocator for multiple GPUs

[makortel]

The CachingHostAllocator uses an associated CUDA stream and device for "asynchronous free" (to support the creation+transfer of "me" pointer in data formats

cmssw/CUDADataFormats/SiPixelDigi/src/SiPixelDigisCUDA.cc

Lines 17 to 25 in e5291a0

	auto view = cs->make_host_unique<DeviceConstView>(stream);
	view->xx_ = xx_d.get();
	view->yy_ = yy_d.get();
	view->adc_ = adc_d.get();
	view->moduleInd_ = moduleInd_d.get();
	view_d = cs->make_device_unique<DeviceConstView>(stream);
	cudaCheck(cudaMemcpyAsync(view_d.get(), view.get(), sizeof(DeviceConstView), cudaMemcpyDefault, stream.id()));
	}

). The implementation missed one detail regarding multiple GPUs: when claiming a previously-cached memory block, the current device may differ from the device of the previous allocation, and in that case, the CUDA event must be re-created for the new device.

This PR fixes that behavior, and should fix the crashes reported in #208 (comment).

@fwyzard