Migrate 200 TB of media and feed it to training

Move about 200 terabytes of media files to new storage and build the input pipeline that streams them to model training. Cover how you copy that much data reliably (parallelism, checksums, resuming after failures, cut-over), the format and sharding you store it in for training, and how the training input pipeline reads, decodes and shuffles it fast enough to keep accelerators busy.