r2.3 cherry-pick request: Shift padded NMS compat window forward to fix TFLite conversion

MLIR bridge doesn't support tf.NonMaxSuppressionV4 legalization that is
conditionally generated by non_max_suppression_padded function.

PiperOrigin-RevId: 320197235
Change-Id: If7242133254680b366771ced50de074ed6180563

Updated estimator version

Add SaveableObjects to SavedModel.

PiperOrigin-RevId: 319856402
Change-Id: I60eeb4e411609511046cc4676c2f3ad630b3769d

Cherry-pick dlpack fix #40843 into r2.3

[CherryPick:r2.3] Support dynamic outputs for XLA on demand ops.

When objects are loaded from the SavedModel, they don't retain their `_gather_saveables_for_checkpoint` functions, which can result in values not being loaded from the checkpoint.

This CL adds a field in the SavedModel proto that stores a save and restore function for each SaveableObject in each node. When loading into Python, the SaveableObjects are restored using the functions.

PiperOrigin-RevId: 318549786
Change-Id: I688c72d7658e1bca98abf373a13a0e15a7fb83e2

PiperOrigin-RevId: 317902879
Change-Id: I6b6dfa54855d5996ac15d4b5c48a5db5dc230025

[CherryPick:r2.3] Fix docker for numpy

[CherryPick:r2.3] Fix the api docstrings for `on_*_batch_begin` methods in `callbacks.Callback` class.

[CherryPick:r2.3] Allow tf.distribute.TPUStrategy to be used with TPUEmbedding API and ensure that LossScaleOptimizer properly rejects it.

[CherryPick r2.3] Fix a critical breakage in `training` argument default value in inference for layers with a default of `training=True` called in e.g. a Sequential container.

I forgot about the trouble using '<' in pip dependencies on the CLI.
This time I verified the build works.

PiperOrigin-RevId: 318332763
Change-Id: I347aee8121464232222e72d89409c600159ee80c

See #40688.

PiperOrigin-RevId: 318122157
Change-Id: Ief46c5610f3aaf0cdd7d43ce1a10d6d87e8e8e01

class.

PiperOrigin-RevId: 318203984
Change-Id: I85e2aa5d4c498b67f3e9130fa45a38173f84b35b

Allow tf.distribute.TPUStrategy to be used with TPUEmbedding API and ensure that LossScaleOptimizer properly rejects it.

PiperOrigin-RevId: 318186211
Change-Id: Id3b9cb8288e5d28ddbaec97d5b35627ab35bc08d

layers with a default of `training=True` called in e.g. a Sequential container.

PiperOrigin-RevId: 318145694
Change-Id: I1af5286824e3a45e1a7d1b8a4fadd7ec223895dc

[CherryPick 2.3] change the size of input to remedy OOM issue.

PiperOrigin-RevId: 317995769
Change-Id: I1358449e989a41c5621e6a4d56e603387be0490d

[CherryPick r2.3] Add the "--define=no_tensorflow_py_deps=true" flag for the windows cp…

builds.

PiperOrigin-RevId: 317968971
Change-Id: I7d4db21474d85620928f3a5ffb1e4cfebaa2be9f

[CherryPick 2.3] Skip data loading error in multi_worker_tutorial_test (the test does not aim to cover this).

Cherry-pick sqlite version bump

PiperOrigin-RevId: 317797271
Change-Id: I8336d7ffeda0836beef0a2d04e633614a44e7fa4

PiperOrigin-RevId: 317934381
Change-Id: I95cdf789f7f5a89d75d45a1b6d67f1ad993cafab

The mapping of the dict input tensors was not correct since it was still using the tensor name, rather than the key of the tensor when build the model. This cause the issue down the stream when the inputs are provided with unknown keys.

We had some backup logic, which will probably do correct things, eg just flat the dict to keep the original order, which was correct most of the case, but not very reliable. In this change, we make the behavior change:

1. When model is build with dict input tensors, the key of the tensor, instead of the name, will be used to map the tensor with input data.
2. Unknown keys in the input data will result into a warning. We didn't throw error since user might do it intentionally, eg using part of the model to test with full input data.

PiperOrigin-RevId: 317776370
Change-Id: I91983443f2b770cb0b45ddb7726f52708cb91d61

PiperOrigin-RevId: 317774876
Change-Id: I6a832236377d403b1ebd24ecbc26025c37dc1c13

PiperOrigin-RevId: 317771299
Change-Id: I71b5c585cef7bd7ef80e66b75e30287fddcf89e2

PiperOrigin-RevId: 317771125
Change-Id: I719c46d97ae1c68ac59dcd1cf8f65d067ddc7658

* Added new API hooks for all the OPs currently supported in TFLM.
* These new APIs still need to be implemented with operator specific parse
  functions but this change allows us to remove the old API and incrementally
  update the implementations.

PiperOrigin-RevId: 317770205
Change-Id: Idaaa687401f2bac5fbf9925e27c04bf536b154ea

- MaybePinSmallOpsToCpu
- MaybePinToResourceDevice
- MaybePinToCustomDevice

We are going to reuse MaybePinSmallOpsToCpu in TFRT but not the other two. Because TFRT doesn't have native Resource neither Custom Device.

PiperOrigin-RevId: 317766813
Change-Id: I43241b5786120ddf39dc4bfff6071239afdfd785