[Rajan-vish]

Fixes #1624

Summary of changes

When computing scores for requested query-passage pairs in M3Embedder.compute_score_single_device, the embedder previously calculated the full cross-matrix (B x B) for dense, sparse, and ColBERT representations before taking the diagonal. For ColBERT especially, constructing the full B x B token interaction tensor scales memory as O(B^2 * L_q * L_p * d), causing unnecessary OOM errors on larger batches.

This PR:

• Adds compute_dense_score_pairwise, compute_sparse_score_pairwise, and compute_colbert_score_pairwise in EncoderOnlyEmbedderM3Model. The ColBERT pairwise calculation uses orch.einsum('bin,bjn->bij', q_reps, p_reps) directly on corresponding pairs, reducing intermediate tensor memory to O(B * L_q * L_p * d) (a B-fold reduction).
• Updates M3Embedder.compute_score_single_device to use these pairwise scoring methods directly when available.
• Adds unit tests in ests/test_m3_pairwise_scoring.py confirming exact numerical alignment between pairwise scoring and the diagonal of full all-pairs scoring.