Implement 1NN Embeddings and Forward Pass

Implement a small machine learning inference pipeline. Part A: Implement batch 1-nearest-neighbor classification. Given a training feature matrix `X_train` with shape `(N, D)`, training labels `y_train` with shape `(N,)`, and a query feature matrix `X_query` with shape `(M, D)`, compute the squared Euclidean distance from each query row to each training row and return one predicted label per query. If there is a distance tie, choose the smallest training index. State the shape of each intermediate tensor. Part B: Implement a dense neural network forward pass. Given input `X` with shape `(B, D0)` and a list of layers `[(W1, b1), ..., (WL, bL)]`, where `Wi` has shape `(D_{i-1}, D_i)` and `bi` has shape `(D_i,)`, apply `XW + b` at each layer. Apply ReLU after every hidden layer and no activation after the final layer. Return the final output and, if useful, the intermediate activations. Validate all dimensions. Part C: Combine the two parts. Transform both training inputs and query inputs through the same neural network to produce embeddings, then perform 1-nearest-neighbor classification in that embedding space.