Add code to train a simple turret controller

This uses Soft Actor-Critic with automatic temperature adjustment to
train a controller for last year's turret.  This is a stepping step to
training a more complicated turret for the swerve.  I need to still
parallelize it, play with the hyper-parameters, and teach it how to go
to a nonzero goal.

Change-Id: I1357b5fbf8549acac4ee0b94ef8f2636867c28ad
Signed-off-by: Austin Schuh <austin.linux@gmail.com>
diff --git a/tools/python/runtime_binary.sh b/tools/python/runtime_binary.sh
index 8498408..95283e0 100755
--- a/tools/python/runtime_binary.sh
+++ b/tools/python/runtime_binary.sh
@@ -27,6 +27,9 @@
     LD_LIBRARY_PATH+=":${path}/../gtk_runtime/lib/x86_64-linux-gnu"
     LD_LIBRARY_PATH+=":${path}/../gtk_runtime/usr/lib/x86_64-linux-gnu"
     LD_LIBRARY_PATH+=":${path}/../gtk_runtime/usr/lib"
+    if [[ -e "${path}/../pip_deps_nvidia_nccl_cu12" ]]; then
+      LD_LIBRARY_PATH+=":${path}/../pip_deps_nvidia_nccl_cu12/site-packages/nvidia/nccl/lib/"
+    fi
     export LD_LIBRARY_PATH
     break
   fi