Reduce peak memory in mid_cnn fusion and load checkpoints to CPU first

[dongyang0122]

Summary

Two small, independent memory fixes in contrastive-pretraining/mr_rate/mr_rate/mr_rate.py:

1. mid_cnn fusion mode (_encode_visual_tokens)

   • Before: all r volumes are batched as (b*r, c, d, h, w) and passed through forward_cnn in a single checkpointed call.
   • After: iterate over the r volumes, calling forward_cnn per volume with run_checkpoint, then torch.stack the per-volume features.
   • Effect: peak CNN activation memory scales with one volume instead of r, mirroring the existing per-volume pattern in late / late_attn. No change to numerical result — same features, just materialized sequentially.

2. MRRATE.load()

   • torch.load(str(path)) → torch.load(str(path), map_location="cpu").
   • Prevents OOM when loading a checkpoint whose tensors were saved on a different / larger GPU layout (e.g., multi-GPU training → fewer-GPU finetune or eval).

Notes

• Only contrastive-pretraining/mr_rate/mr_rate/mr_rate.py is touched; no changes to public API, CLI, or checkpoint format.
• Matches the repo's existing style of sequential per-volume processing with gradient checkpointing.