Summary

When gather_across_devices=True, GISTEmbedLoss and CachedGISTEmbedLoss build the
positive-protection mask without the cross-device offset, so on rank > 0 the wrong columns are
protected and true positives can be suppressed.

In both losses the CE targets and the guided-similarity diagonal already use the rank offset (the
cached loss adds begin for its minibatch loop):

range_labels = torch.arange(offset, offset + batch_size, ...)        # CE targets
guided_sim = guided_ap_sim.diagonal(offset=offset).view(-1, 1)       # positive threshold (GISTEmbedLoss)
guided_sim = guided_ap_sim.diagonal(offset=offset + begin).view(-1, 1)  # CachedGISTEmbedLoss

but the positive mask did not:

# GISTEmbedLoss
positive_mask = torch.eye(*guided_ap_sim.shape, ...)                 # protects columns [0 .. batch-1]
# CachedGISTEmbedLoss
positive_mask = torch.eye(*guided_ap_sim.shape, ...).roll(begin)

On rank 1 with world_size=2 and batch_size=3, the CE target for local row r is column
offset + r (i.e. [3, 4, 5]), but the mask protects [0, 1, 2]. With margin > 0 the true
positive itself exceeds positive_sim - margin, so the absolute-margin false-negative suppression
sets the target logit to -inf and cross-entropy then returns +inf. This can be hidden in rank-0
or single-process runs because offset = 0.

gather_across_devices was added to these losses in #3442, and #3453 patched the offset alignment
for the in-batch-negative losses (MultipleNegativesRankingLoss and the cached / symmetric
variants), which select their identity mask by global index (identity[local_batch] with
local_indices = arange(offset, ...)). The two GIST losses were not part of that patch and kept the
offset-unaware mask.

This builds the mask with the same global index as the CE targets:

positive_mask = torch.zeros_like(guided_ap_sim, dtype=torch.bool)
rows = torch.arange(guided_ap_sim.size(0), ...)
positive_mask[rows, offset + rows] = True            # + begin in the cached minibatch loop

When offset == 0 this is identical to the previous mask, so rank 0 / single-GPU behavior is
unchanged.

Tests

Added tests/sentence_transformer/losses/test_gist_embed.py and
tests/sentence_transformer/losses/test_cached_gist_embed.py. They simulate rank 1 / world 2 in a
single process (monkeypatching the gather and rank helpers; the cached test parametrizes
mini_batch_size so begin = 0 and begin > 0 are both covered) and assert that with margin > 0
the loss stays finite and that the protected column equals the CE target. Both regression tests fail
on the current code and pass with the fix.

pytest tests/sentence_transformer/losses/test_gist_embed.py tests/sentence_transformer/losses/test_cached_gist_embed.py