Shuffle before batching in TF load_data_fashion_mnist

[Chessing234]

Bug

`load_data_fashion_mnist` in the TensorFlow build of `d2l` (`chapter_appendix-tools-for-deep-learning/utils.md` → `d2l/tensorflow.py`) chains the train pipeline as `.batch(batch_size).shuffle(len(mnist_train[0]))`:

```python
return (
tf.data.Dataset.from_tensor_slices(process(*mnist_train)).batch(
batch_size).shuffle(len(mnist_train[0])).map(resize_fn),
tf.data.Dataset.from_tensor_slices(process(*mnist_test)).batch(
batch_size).map(resize_fn))
```

Because `tf.data.Dataset` operators fire top-down, this first groups items into fixed `batch_size` blocks and only then shuffles those blocks. Batch composition is frozen for the lifetime of the dataset — every epoch sees the exact same 64-item groups, only their order changes. This defeats the item-level stochasticity SGD relies on.

Root cause

The two transform calls are reversed. Everywhere else in the same module the idiomatic order is used, e.g. `DataModule.get_tensorloader` in `d2l/tensorflow.py` (`chapter_builders-guide/oo-design.md`):

```python
tf.data.Dataset.from_tensor_slices(tensors).shuffle(
buffer_size=shuffle_buffer).batch(self.batch_size)
```

That's `.shuffle(...).batch(...)` — shuffle at item granularity, then batch from the shuffled stream. The Fashion-MNIST loader is the single place that got the two composed in the wrong direction. The PyTorch counterpart (`torch.utils.data.DataLoader(..., shuffle=True)`) and the MXNet counterpart (`gluon.data.DataLoader(..., shuffle=True)`) both correctly shuffle at item granularity.

Why the fix is correct

Swapping to `.shuffle(len(mnist_train[0])).batch(batch_size).map(resize_fn)` fills a shuffle buffer of size 60000 (the full training set) with individual items before batching, so each epoch produces a fresh, fully randomized set of batches — matching the torch/mxnet loaders and the rest of the `tf.data` pipelines in this package. The test pipeline is left untouched (no shuffle), and the `map(resize_fn)` stage still runs on the batched tensors exactly as before.

The fix is applied in both the `utils.md` source (canonical) and the generated `d2l/tensorflow.py` so the change is visible to readers before `d2lbook build lib` is re-run.