Don't import TensorFlow when defining an Audio feature

[raashish1601]

Fixes #11233.

_AudioDecoder.__init__ called tf.dtypes.as_dtype right away, so just defining a tfds.features.Audio (for example in a builder's _info()) imported TensorFlow. That breaks TensorFlow-free use such as tfds.data_source on audio datasets.

This turns _dtype into a property, so TensorFlow is only used when a decoder actually needs the TF dtype.

Tests: added test_construction_does_not_use_tensorflow. I couldn't run the TF test files locally (TensorFlow doesn't import on native Windows here), but in an environment without TensorFlow tfds.features.Audio(sample_rate=16000) raises ModuleNotFoundError before this change and works after it.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.