GPUDirect to local SSD

[Fizzbb]

Issues to investigate

1. steps to install/enable a NVMe SSD (not plugin and directly visible under Ubuntu)
2. steps/configurations to make SSD visible to GPU (GPU model requirements)
3. cuda level code changes to make GPU read/write direct from/to SSD
4. python level code changes
5. simple mnist training job completion time comparison before and after

Time Saving Theoretically calculation (?)
E.g. 150MB cifar10 data, CPU read time (from SSD to CPU memory) xx sec (a), GPU copy time fromHtoD xx sec (b). GPU direct SSD read time xx sec (c). read time is calculated by 150MB/theoretical bandwidth
time saving: a + b - c