Support warm-start in tfra.dynamic_embedding

[dakabang]

Description

In tensorflow we use tf.train.Checkpoint(tf2) or tf.compat.v1.train.Saver(tf1) for saving/restoring model, but I found it a little hard when I use these APIs to restore part of the model in tfra to perform transfer learning. In recommendation system, it's a common trick to speed up model convergence using embedding warmup, and the dense part of the model is trained from scratch, hence partial restore. The demonstrate the API design for partial restore are as follows.

restore_op = tfra.dynamic_embedding.warm_start(ckpt_to_initialize_from='/path/to/ckpt/', vars_to_warm_start='.*embeddings.*')
with tf.compat.v1.Session() as sess:
   sess.run(restore_op)
   ...

Changes

• Add tfra.dynamic_embedding.warm_start to support warm-start in tfra model.

[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).

For more information, open the CLA check for this pull request.

[rhdong]

Hi @dakabang , please sign the CLA by following the guide, thank you!
For more information, open the CLA check for this pull request.

❌ 18f7c4f PR Opener: @dakabang <zhang​@yeah.net>
❌ 18f7c4f Author: <eth*han​@vipshop.com>

[rhdong]

recommend using different name for testing different name function in one case, or you'd better to write another case.