[fix] Write the sparsity CSV columns once in three sparse evaluators - #3985

SparseBinaryClassificationEvaluator, SparseEmbeddingSimilarityEvaluator and SparseTripletEvaluator write a CSV whose header has more columns than its rows, so every value ends up under the wrong column name. It is the same bug as #3512, which you fixed for SparseInformationRetrievalEvaluator in #3456 but which is still there in these three.

Each of them overrides _append_csv_headers to append its active_dims / sparsity_ratio columns (sparse_binary_classification.py:136, sparse_embedding_similarity.py:114, sparse_triplet.py:129). The dense superclass calls that hook twice whenever similarity_fn_names is left unset, which is the default: once from __init__ with an empty list, and again from __call__ once model.similarity_fn_name has filled it in (binary_classification.py:121 and :180, embedding_similarity.py:132 and :187, triplet.py:157 and :196). The metric columns are only added on the second call, but the sparse columns were appended unconditionally, so they landed in csv_headers both times.

The row does not grow to match, because the metric values come from the scores dict and the sparsity values are appended once by append_to_last_row. SparseTripletEvaluator on main writes a 15 column header over a 9 column row:

epoch,steps,anchor_active_dims,anchor_sparsity_ratio,positive_active_dims,positive_sparsity_ratio,negative_active_dims,negative_sparsity_ratio,accuracy_dot,anchor_active_dims,anchor_sparsity_ratio,positive_active_dims,positive_sparsity_ratio,negative_active_dims,negative_sparsity_ratio
-1,-1,0.5,77.25,0.9974690387261648,102.75,0.9966335757814035,102.75,0.9966335757814035

The fix is the guard SparseInformationRetrievalEvaluator._append_csv_headers and SparseNanoBEIREvaluator._append_csv_headers already use: append the sparse columns only when the superclass is adding metric columns too. Passing similarity_fn_names explicitly was already fine and stays fine, since then only the __init__ call adds anything.

The new test builds every SparseEncoder evaluator that writes a CSV and checks the header has no repeated column and exactly as many cells as the row.

$ pytest tests/sparse_encoder/test_evaluators.py -q
3 failed, 4 passed        # on main: binary_classification, embedding_similarity, triplet
7 passed                  # with this change
$ pytest tests/sparse_encoder tests/sentence_transformer/evaluation -q
442 passed, 2 skipped in 306s
$ pre-commit run --files <the four files touched>
ruff check Passed, ruff format Passed, typos Passed

I also checked the CSVs the other sparse evaluators write, before and after: SparseInformationRetrievalEvaluator, SparseNanoBEIREvaluator, SparseRerankingEvaluator, SparseTranslationEvaluator and SparseMSEEvaluator are unaffected, and repeated calls appending to an existing file stay aligned.