fix(cache): clone embeddings before save to avoid persisting whole ba…

[JiaqiL10]

Fixes #12

Compatibility / Impact
Non-breaking. Existing caches still load correctly (values are unchanged); no migration is required.
After the fix, newly built caches shrink from batch_size × dim to dim per file (e.g. 16× smaller at --batch-size 16, 128× at --batch-size 128). Old caches can optionally be rebuilt to reclaim disk.
Embedding values are bit-for-bit identical — no effect on any already-trained or in-progress models.
Train-time side benefit: torch.load no longer pulls the entire [batch, dim] storage into memory per sample, reducing DataLoader memory and disk I/O.
Verification
Build a cache before and after the fix from the same input and assert:

Values unchanged — loaded vectors are bitwise-equal.
Size fixed — per-file size is on the order of dim × 4 bytes and no longer scales with batch_size.
import torch, os

emb = torch.load(path, map_location="cpu")["all_findings"] # dict value

1) values: identical to a freshly-cloned compute (no change in numbers)

assert torch.equal(emb, emb.clone())

2) size: file no longer carries the whole batch storage

real = emb.numel() * emb.element_size() # dim * 4B (e.g. 16 KB)
print("file:", os.path.getsize(path), " real tensor:", real)
print("storage elems:", emb.untyped_storage().nbytes() // emb.element_size()) # == dim, not batchdim
print("storage_offset:", emb.storage_offset()) # 0 after fix (was idim)
Before the fix: file ≈ batch_size × dim × 4B, storage elems == batch_size × dim, storage_offset == i × dim.
After the fix: file ≈ dim × 4B, storage elems == dim, storage_offset == 0.