Seq2Seq inference wrapper as an NNModuleSpec

The Seq2Seq spec model defines encoder/decoder math and differentiable training helpers. This file provides a small inference-oriented NNModuleSpec wrapper so it can be composed/exported.

def Spec.Seq2SeqModuleSpec {α : Type} [Context α] {srcVocabSize tgtVocabSize embedDim hiddenDim srcSeqLen tgtSeqLen : ℕ} (m : Seq2SeqSpec α srcVocabSize tgtVocabSize embedDim hiddenDim) (start_token : ℕ) (_h1 : srcVocabSize ≠ 0) (h2 : tgtVocabSize ≠ 0) (_h3 : embedDim ≠ 0) (_h4 : hiddenDim ≠ 0) (_h5 : srcSeqLen ≠ 0) (h6 : tgtSeqLen ≠ 0) : ModSpec.NNModuleSpec α (Shape.dim srcSeqLen (Shape.dim srcVocabSize Shape.scalar)) (Shape.dim tgtSeqLen (Shape.dim tgtVocabSize Shape.scalar))

Seq2Seq inference module wrapper (one-hot input, greedy decoding).