chore: rebase off of main + resolve conflicts + regenrate

Signed-off-by: anubhutiv <anubhutiv@nvidia.com>

Author: anubhutivyas

docker/rl/Dockerfile.nmp-rl-base

@@ -243,13 +243,6 @@ RUN uv pip install --python /opt/nemo_rl_venv/bin/python --no-deps -e /opt/nemo-
 #
 # NOT prefetched: automodel, mcore, sglang, trtllm, modelopt-quant
 # and the async-GRPO bookkeeping actors; those build on the node if a config selects them.
-#
-# automodel is on the GRPO + LoRA path: LoRA is implemented only in DTensorPolicyWorkerV2, the
-# compiler selects it via policy.dtensor_cfg._v2, and the registry maps that actor to
-# PY_EXECUTABLES.AUTOMODEL. Because it is not prefetched, a LoRA run builds that venv on the node
-# at job start - slow at best, and a hard failure on a deny-egress cluster. Add
-# dtensor_policy_worker_v2.DTensorPolicyWorkerV2 to the filters and the verification loop below to
-# fix that. Full-weight GRPO stays on V1 -> fsdp, which is prefetched.
 RUN <<"EOF" bash -exu
 export UV_LINK_MODE=symlink
 uv run nemo_rl/utils/prefetch_venvs.py \