Revert "Revert "Add Recurrent Layers""

[dan-zheng]

Reverts #94.

---

The major blocking crash is fixed in swiftlang/swift#24114.
@differentiable must also be removed from var zeroState to prevent another error/crash.

[dan-zheng]

Confirmed that this test passes (adapted from Tests/DeepLearningTests/LayerTests.swift):

import TensorFlow

let weight = Tensor<Float>(ones: [7, 5]) * Tensor<Float>([0.3333, 1, 0.3333, 1, 0.3333])
let bias = Tensor<Float>(ones: [5])
var cell = SimpleRNNCell<Float>(inputSize: 2, hiddenSize: 5)
cell.weight = weight
cell.bias = bias
let state = Tensor<Float>(ones: [1, 5]) * Tensor<Float>([1, 0.2, 0.5, 2, 0.6])
let input = Tensor<Float>(ones: [1, 2]) * Tensor<Float>([0.3, 0.7])
let output = cell.applied(to: input, state: state).state
let expected = Tensor<Float>([[2.76649, 6.2999997, 2.76649, 6.2999997, 2.76649]])
assert(output == expected)