Implement a custom FSDP strategy for benchmarking loads from boot disk

[abhibyreddi]

This is needed to be able to benchmark checkpoint saves/loads to/from boot disk in distributed environment

• When run with --save_only, lightning's FSDP strategy is used. Load calls are skipped.
• When run with --load_only newly added LoadFromBootDiskFSDP is used which saves with dataflux. All nodes then copy all the contents of the GCS bucket to boot disk and then load their checkpoints. Avg. time to save one checkpoint will be reported as skipped.
• Move all custom strategy classes to demo/lightning/checkpoint/multinode/strategies.py

• Tests pass - ran manually on a VM with GPUs and on a GKE cluster with 2 nodes
• Appropriate changes to documentation are included in the PR - follow up PR will update the README

[abhibyreddi]

Contents of demo/lightning/checkpoint/multinode/strategies.py are mostly reviewed already except for the CustomFSDPStrategy class which is introduced in this PR.