ImageNet: release configurations

[yaroslavvb]

current best-known list of configurations to be used for release @bearpelican

• 1 machine: xar_args_pytorch
• 4 machines: x4ar_args_bench
• 8 machines: x8ar_args_benchmark
• 16 machines: x16ar_args_benchmark
• 32 machines: x32ar_throughput

[yaroslavvb]

x4ar_args_bench is broken because of missing sz/352 folder (it's missing on EFS, so it didn't make it into snapshots used by replicate_imagenet.py), so instead

• 4 machines: x4ar_args

[bearpelican]

@yaroslavvb
I'll do some testing on this to see the performance difference. Perhaps it's not worth it to have that folder

The current pytorch_source ami has all the necessary resize folders. I can copy those over to our snapshots.

[yaroslavvb]

Dropping the 32 machine configuration for release, testing 1, 4, 8 and 16 on new account