Upgrade to LightGBM 3.3.1 (Kaggle#1106)

http://b/207851889

Upgrade to LightGBM 3.3.1 (Kaggle#1106)

http://b/207851889

Remove /usr/local/cuda/compat from LD_LIBRARY_PATH (Kaggle#1092)

* Remove /usr/local/cuda/compat from LD_LIBRARY_PATH

http://b/203659565

* Add CUDA stubs to LD_LIBRARY_PATH to allow building on a CPU-only machine

* Add symlink for libcuda.so

Remove /usr/local/cuda/compat from LD_LIBRARY_PATH (Kaggle#1092)

* Remove /usr/local/cuda/compat from LD_LIBRARY_PATH

http://b/203659565

* Add CUDA stubs to LD_LIBRARY_PATH to allow building on a CPU-only machine

* Add symlink for libcuda.so

Add /opt/bin to GPU image PATH env variable (Kaggle#1089)

NVIDIA binaries from the host are mounted to /opt/bin.

This is how `nvidia-smi` is made accessible inside the user session docker container. See attached bugs for additional context.

b/202852505

Add /opt/bin to GPU image PATH env variable (Kaggle#1089)

NVIDIA binaries from the host are mounted to /opt/bin.

This is how `nvidia-smi` is made accessible inside the user session docker container. See attached bugs for additional context.

b/202852505

Add /opt/bin to GPU image PATH env variable (Kaggle#1089)

NVIDIA binaries from the host are mounted to /opt/bin.

This is how `nvidia-smi` is made accessible inside the user session docker container. See attached bugs for additional context.

b/202852505

Add /opt/bin to GPU image PATH env variable (Kaggle#1089)

NVIDIA binaries from the host are mounted to /opt/bin.

This is how `nvidia-smi` is made accessible inside the user session docker container. See attached bugs for additional context.

b/202852505

Prune image from GPU agent disk. (Kaggle#1088)

* Prune image from GPU agent disk.

It was currently only running on the CPU (ephemeral) agents. 
Switched to be run on the GPU agent building the image to prevent build errors caused by disk filling up.

* Reduce prune window

Prune image from GPU agent disk. (Kaggle#1088)

* Prune image from GPU agent disk.

It was currently only running on the CPU (ephemeral) agents. 
Switched to be run on the GPU agent building the image to prevent build errors caused by disk filling up.

* Reduce prune window