[YOLOv5] Change default cache from ram to disk

[dsikka]

For some quick benchmarking, I had two training_aware runs using the following model and data. One was with disk caching and one was with ram. For the first 50 epochs, the training time with disk was about 5 extra minutes (about an 8% slowdown) for disk caching vs ram.

Compared the following on disk caching vs ram for the first 50 epochs:

• Model: zoo:cv/detection/yolov5-n/pytorch/ultralytics/coco/base-none
• Data: VOC.yaml
• Use Case: cv-detection
• 1 NVIDIA RTX A4000 GPU

For the sake of time, I only looked at the time for the first 50 epochs but we could potentially consider a longer running job for additional benchmarking

To go back to ram, you can provide the the cache kwarg and set it to ram

[bfineran]

could you give a rough estimate of performance with disk vs ram for a local run? LGTM after that