Build a full transformer block from scratch: multi-head attention, feed-forward network, residuals, and layer norm.

Compose pre-norm attention and MLP sublayers with residual additions: x←x+Attn(Norm(x)), then x←x+MLP(Norm(x)). Keep dropout and masks explicit and verify batch, head, sequence, and feature dimensions.

Test empty input, one-element input, duplicates, boundary indices, invalid states, and the largest allowed size; state time and space complexity.