fix: prevent singleton shrinkage chunks from modifying covariance

[Tal-Golan]

Summary

Fix an in-place modification in the Torch shrinkage-NLL path that can cause
GSN to select lambda = 0 even when it is not optimal.

I encountered this with a large covariance matrix when the available memory
caused the automatic batching code to process one shrinkage level at a time.

Cause

The code intends to create a writable copy before modifying each chunk:

covs = c_t.unsqueeze(0).expand(chunk_size, N, N).contiguous()
covs.mul_(alphas_chunk[:, None, None])

For chunks containing multiple levels, contiguous() creates a new tensor.
For a singleton chunk, however, the expanded view may already be contiguous,
so contiguous() can return storage shared with c_t.

The following in-place multiplication then modifies c_t. Later shrinkage
levels are evaluated using this modified tensor, which can incorrectly make
the selection fall back to lambda = 0.

Singleton chunks typically occur when the available memory is low relative to
the covariance matrix size.

Commits

This PR contains two commits:

1. test: show that one shrinkage level changes the input covariance

   Adds a focused test that demonstrates the issue using a singleton shrinkage
   level. The test shows that the Torch path modifies the input covariance and
   also compares its NLL with the NumPy implementation. It fails before the
   fix.

2. fix: copy the covariance before changing it

   Replaces expand(...).contiguous() with an explicit contiguous clone:

   covs = c_t.unsqueeze(0).expand(chunk_size, N, N).clone(
       memory_format=torch.contiguous_format
   )

   This guarantees independent writable storage before the in-place
   multiplication. The test introduced by the first commit passes after this
   change.

Validation

• The focused test fails on the first commit and passes on the second.
• Complete suite: 145 passed, 2 skipped.

[kendrickkay]

Thanks Tal. @jacob-prince is busy at the moment but will get to this next week.

[jacob-prince]

yes thanks @Tal-Golan for raising this! i'll handle asap.