Struct std::DeviceStats

struct DeviceStats {
    unsigned long launches;
    unsigned long gemms;
    unsigned long uploads;
    unsigned long downloads;
    unsigned long uploadBytes;
    unsigned long downloadBytes;
    unsigned long cpuLaunches;
}

Counters since the program started (or the last device_stats_reset): what moved between host and device, and how many kernels and GEMMs ran. A training step that keeps everything on the device shows no downloads besides the loss it reads, and no uploads besides its batch. cpuLaunches counts the kernels and GEMMs a backend ran on the CPU (device_ref.sc) for want of a GPU version: 0 when the GPU ran them all.