Inconsistent CLI parameters to specify resources requests and limits across training and inference

[giuseppeporcelli]

For training, the CLI parameters are:

  --accelerators INTEGER          Number of accelerators (GPUs/TPUs)
  --vcpu TEXT                     Number of vCPUs
  --memory TEXT                   Amount of memory in GiB
  --accelerators-limit INTEGER    Limit for the number of accelerators (GPUs/TPUs)
  --vcpu-limit TEXT               Limit for the number of vCPUs
  --memory-limit TEXT             Limit for the amount of memory in GiB

For inference (custom endpoints), the CLI parameters are:

  --resources-requests JSON       JSON object of resource requests, e.g. '{"cpu":"1","memory":"2Gi"}'
  --resources-limits JSON         JSON object of resource limits, e.g. '{"cpu":"2","memory":"4Gi"}'

Also, as mentioned in issue 306 (#306), the CLI should offer a consistent way to specify EFA (and Neuron) devices.