Moved code to lightning

Updating codebase to make use of pytorch-lightning

• Main train/valid loops are handled by lightning
• Removed device-handling code
• Each model implements its own train/valid/predict steps
• Updated config files for use with lightning
• Using lightning checkpointing and logging callbacks
• Updated DeviceMonitor to be compatible with lightning

src/ptbench/engine/trainer.py

@@ -42,6 +42,7 @@ class AcceleratorProcessor:
         if len(split_accelerator) > 1:
             devices = split_accelerator[1:]
             devices = [int(i) for i in devices]
+            os.environ["CUDA_VISIBLE_DEVICES"] = devices
         else:
             devices = "auto"