fix MatryoshkaLoss bug: sort sampled dimension indices to maintain descending dimension order (#3203)

When randomly sampling dimension indices for MatryoshkaLoss, ensure they are
processed in ascending order (descending dimension). Previously, when a smaller dimension was be processed
first, the ForwardDecorator would cache the output tensor and then shrink it.
Since the cached tensor and the shrunk output reference the same underlying
memory, the cache would contain truncated embeddings, making them unusable for
subsequent larger dimensions.

Author: emapco

sentence_transformers/losses/MatryoshkaLoss.py

@@ -218,6 +218,7 @@ def forward(self, sentence_features: Iterable[dict[str, Tensor]], labels: Tensor
             dim_indices = range(len(self.matryoshka_dims))
             if self.n_dims_per_step > 0 and self.n_dims_per_step < len(dim_indices):
                 dim_indices = random.sample(dim_indices, self.n_dims_per_step)
+                dim_indices.sort()
 
             loss = 0.0
             for idx in dim_indices: