Fix 3D attn_mask head/batch layout in T2V/V2T cross-attention (repeat → repeat_interleave)

[komatta-san]

Problem

nn.MultiheadAttention expects a 3D attn_mask of shape (N * num_heads, L, S) indexed as batch_idx * num_heads + head_idx. The mask is built with .repeat(nhead, 1, 1), which tiles instead: [M0, M1, M0, M1] rather than the expected [M0, M0, M1, M1]. Heads then attend using the wrong sample's padding mask.

Impact

When num_heads > 1 and samples have different valid lengths (the normal case), a sample's output depends on the other samples in the batch, breaking determinism. At batch_size == 1 the two ops are identical, so single-sample outputs are unchanged.

Fix

Use .repeat_interleave(self.nhead, dim=0) in all affected encoder layers: CIM.py (T2V+V2T), cg_detr_, qd_detr_, taskweave_, tr_detr_transformer.py.

[awkrail]

lgtm. @h-munakata Could you merge it?