build: bump TensorFlow to 2.3.0

[compnerd]

The TensorFlow 2.3.0 release is GA'ed. Update the version that we use.

[texasmichelle]

I believe that setting reshuffleForEachIterator explicitly to false will resolve the test failure:

let shuffled = dataset.shuffled(sampleCount: 5, randomSeed: 42, reshuffleForEachIterator: false)

[texasmichelle]

The failure is caused by shuffle behaving differently, which I suspect is caused by an upstream bugfix. Datasets.shuffled() defaults to reshuffleForEachIterator: true, which did not have any effect before.

Here's a reduction:

import TensorFlow

let scalars = Tensor<Float>(rangeFrom: 0, to: 5, stride: 1)
let dataset = Dataset(elements: scalars)
print("dataset: \(dataset.flatMap { $0.scalars })")
let shuffled = dataset.shuffled(sampleCount: 5, randomSeed: 42)
print("shuffled: \(shuffled.flatMap { $0.scalars })")
print("shuffled: \(shuffled.flatMap { $0.scalars })")
print("shuffled: \(shuffled.flatMap { $0.scalars })")

Nightlies:

dataset: [0.0, 1.0, 2.0, 3.0, 4.0]
shuffled: [0.0, 4.0, 1.0, 3.0, 2.0]
shuffled: [0.0, 4.0, 1.0, 3.0, 2.0]
shuffled: [0.0, 4.0, 1.0, 3.0, 2.0]

TF 2.3:

dataset: [0.0, 1.0, 2.0, 3.0, 4.0]
shuffled: [0.0, 4.0, 1.0, 3.0, 2.0]
shuffled: [3.0, 4.0, 1.0, 0.0, 2.0]
shuffled: [2.0, 4.0, 3.0, 0.0, 1.0]

Passing in reshuffleForEachIterator: false gives us behavior that matches TF 2.2:

dataset: [0.0, 1.0, 2.0, 3.0, 4.0]
shuffled: [2.0, 1.0, 3.0, 4.0, 0.0]
shuffled: [2.0, 1.0, 3.0, 4.0, 0.0]
shuffled: [2.0, 1.0, 3.0, 4.0, 0.0]

[texasmichelle]

I wonder why the shuffled values are different, given the same seed? It's probably fine just to update the expected values in the test.