[Train / Release] Fix Huggingface TRL release test (ray-project#63188)

## Description
The HuggingFace Transformer Reinforcement Learning release test is
failing from `(RayTrainWorker pid=3665, ip=10.0.159.246) ERROR: Flag
‘grpc_experiments' was defined more than once but with differing types.
Defined in files
'external/com_github_grpc_grpc/src/core/config/config_vars.cc' and
'workspace/grpc/src/core/config/config_vars.cc'. [repeated 3x across
cluster]`
This appears to be due to a crash in ray's grpc and tensorflow's grpc
that HF transformers will import.

To address this issue, this PR we just uninstall tensorflow as it isn't
used in the release script

---------

Signed-off-by: Mark Towers <mark@anyscale.com>
Co-authored-by: Mark Towers <mark@anyscale.com>

Author: 2 people

release/ray_release/byod/byod_huggingface_trl.sh

@@ -5,9 +5,9 @@
 set -exo pipefail
 
 # Install TRL and math_verify
-pip3 install --no-cache-dir "trl[vllm]" math_verify
-pip3 install --no-cache-dir --force-reinstall numpy pandas tf_keras
-# vllm requires numpy>=2.x; upgrade tensorflow and matplotlib to a version compatible with numpy 2.x.
-pip3 install --no-cache-dir --upgrade tensorflow matplotlib
-# wandb and comet were installed but didn't have API keys which raised errors and warnings
-pip3 uninstall -y wandb comet_ml
+pip3 install --no-cache-dir "trl[vllm]==1.0.0" math_verify
+pip3 install --no-cache-dir --force-reinstall numpy pandas
+# vllm requires numpy>=2.x; upgrade matplotlib to a version compatible with numpy 2.x.
+pip3 install --no-cache-dir --upgrade matplotlib
+# `transformers` auto-imports TF when present and TF's bundled grpc collides with Ray's grpc on the `grpc_experiments` absl flag
+pip3 uninstall -y wandb comet_ml tensorflow tf_keras