Update version numbers for TensorFlow 2.4.0-rc3

[CherryPick:r2.4] Rename exec_tools to tools

[CherryPick:r2.4]Change NCCL_PATCH to 6

Follow up to #43156
Based on
https://github.com/bazelbuild/bazel/issues/12059#issuecomment-725641997
exec_tools might no longer be needed and hence can be replaced by tools.
This fixes various build failures caused by missing environment
variables in environments where they are required, e.g. using custom
compilers.

[r2.4 cherry-pick] Include C logging API into part of the libtensorflow_framework.so

[CherryPick:r2.4][tf.data] Making sure setting tf.data threading options does not prevent auto-sharding from using file-level sharding.

[CherryPick:r2.4] add per_replica support for keras

We're building against NCCL 2.7.6, but were incorrectly reporting the version as
2.7.3.

PiperOrigin-RevId: 343586102
Change-Id: I4cf2c4f1b7fdd97ff94662baf9c586b1055681ee

[tf.data] Making sure setting tf.data threading options does not prevent auto-sharding from using file-level sharding.

PiperOrigin-RevId: 343746154
Change-Id: I241e0c951c46086e9f043588ddb3aa0eff23e593

This PR includes C logging API into part of the libtensorflow_framework.so
as otherwise the file system plugins will not be able to link to TF_Log
to send logs to tensorflow logging system.
Signed-off-by: NYong Tang <yong.tang.github@outlook.com>

[CherryPick:2.4]Do not use fused batch norm in the 5D case.

[Cherrypick:r2.4]Disables //tensorflow/python/kernel_tests/boosted_trees:[training_ops_test,prediction_ops_test,resource_ops_test] on Mac OS because of segfaults

Update version numbers for TensorFlow 2.4.0-rc2

[CherryPick:r2.4]Suppress warnings when listing functions for serialization.

Always enable get_next_as_optional unless the dataset is finite.

[CherryPick:r2.4]Fix single pip package renaming bug

Don't enable get_next_as_optional if the dataset is finite.

[CherryPick:r2.4] Bump libjpeg-turbo from 2.0.4 to 2.0.5

2.4.0 cherry-pick request: update tensorboard dependency to 2.4.x

[CherryPick:r2.4] Expose Logging C API in pip package

Add to release notes: deterministic tf.image.resize (bilinear)

PiperOrigin-RevId: 342271395
Change-Id: Ieedeeb7817f794c41dae3eb06b9b0d3b607fa89a

There is a bug in the single pip package renaming code: we do a replacement over the entire contents of `METADATA` and this causes the `tensorflow_estimator` dependency to be replaced with `tensorflow_gpu_estimator` (on 1.15 it was `tensorflow_cpu_estimator`). These packages don't exist by themseleves, Estimator has no CPU/GPU split. Previously this required a manual alteration of the Estimator package to fake it being the CPU/GPU version and a manual upload for that, but we should move away from this manual step as it always causes issues with new releases.

See for example #44775 (there are a few more similar issues, both internally and externally, but this is the most recent one).

We should build each pip package instead of doing the renaming. We do that on Linux/Mac already but Windows builds take too long so rather than rebuilding we just fake the new package via this renaming function. Future work in this area is needed to get rid of the renaming function, eventually removing it completely from both TF and TF ecosystem packages.

PiperOrigin-RevId: 341915841
Change-Id: I2bd4c3621e581ccf31e7bdd52958937b93971b90

PiperOrigin-RevId: 341945136
Change-Id: I79fdec366be2119b6a28063f193e6cecb7a5f9e2

PiperOrigin-RevId: 341945136
Change-Id: I79fdec366be2119b6a28063f193e6cecb7a5f9e2

It looks like the latest libjpeg-turbo is 2.0.5 so this PR
bumps the version (currently on 2.0.4).
Signed-off-by: NYong Tang <yong.tang.github@outlook.com>