Keep per-query negative groups intact in M3 training

[Afloat16]

Per-query negative groups are currently interpreted as globally concatenated passage groups in two places. With no_in_batch_neg_flag=True, M3 distillation receives (batch_size, group_size) scores but starts each row's targets at row * group_size. Batch size 2 and group size 3 raises IndexError: Target 3 is out of bounds.

The unified M3 ensemble has the same layout problem. Its three score matrices are already local, but they are gathered again. The passage-vector reduction introduced in #1424 makes the inferred group size one: an otherwise valid batch of 2 queries with 3 passages each gets an ensemble contrastive loss of zero, discarding the ensemble negatives. When batch size exceeds group size, it still indexes beyond the score columns. The official m3_same_dataset.sh includes datasets with these local groups.

Keep already-local score groups intact, use zero-based target columns for local M3 distillation, and retain the full passage group when combining the three M3 score branches. Globally concatenated in-batch/cross-device layouts keep their existing target offsets. This follows the conditional likelihood: each query's denominator contains its own remaining passages, and selected target columns are removed within that same group.

Validation: 47 native CPU checks pass; the original implementation fails 23 checks. Tests compare both losses and score gradients against an independent per-query conditional-likelihood calculation in float32/float64. Native tiny BERT encoders exercise dense and unified M3 training, all three heads' gradients and SGD updates, supervised/KL/M3 losses, and self-distillation. Six native in-batch objective controls pass on both versions. The complete FlagEmbedding package imports normally with Transformers 4.57.1 and PyTorch; no source extraction, module placeholders or model mocks are used.

The tests use a local synthetic vocabulary and small randomly initialized models. Pretrained checkpoints, GPU/DDP training and retrieval-quality evaluation were not run. This change does not address the distributed KL loss scaling discussed in #1604.

Related: #1410 and its merged repair #1424; the new regressions reproduce the remaining local-score layout errors on current master.