CUDA: disable timer support on cuda events

[devreal]

According to the docs, disabling timers on events can improve performance.

I wrote a small benchmark to test the difference on xsdk:

    auto now = system_now();
    for (int i = 0; i < NUM_ITER; ++i) {
        ret = cudaEventRecord(event);
        assert(cudaSuccess == ret);
	    do {
           ret = cudaEventQuery(event);
	    } while (cudaErrorNotReady == ret);
        assert(cudaSuccess == ret);
    }

    auto then = system_now();
    std::cout << "Time " << duration_in_mus(now, then) << " us, avg " << duration_in_mus(now, then) / (double)NUM_ITER << " us " << std::endl;

The difference is about an order of magnitude:

Timers enabled: 5.18817
Timers disabled: 0.46593