fix: MPS cache not flushed in alora test GPU cleanup

[planetf1]

PR #765 added CUDA GPU cleanup to the alora integration tests (test_alora_train_integration.py:316-318, 398-400), but the MPS equivalent is missing. On Apple Silicon with MPS-capable PyTorch (>= 2.8), the GPU memory won't be reclaimed between tests.

The fix is straightforward — the pattern already exists in memory_cleaner() (conftest.py:448-449):

if torch.backends.mps.is_available():
    torch.mps.empty_cache()

Just needs adding after the torch.cuda blocks in both cleanup sites.