fix(vllm_inference): pin transformers<5 for vLLM 0.15.x compat

The install_vllm.sh script runs `pip install --upgrade transformers[torch]`
after installing vLLM 0.15.1, which upgrades transformers to 5.x.
vLLM 0.15.x requires transformers<5, and the 5.x upgrade causes silent
embedding mismatches (98% of values differ, max abs diff ~0.054) when
running the ESM2 pooling tests.

Pin transformers to >=4.56.0,<5 to stay within vLLM's compatibility
window.

Signed-off-by: svc-bionemo <267129667+svc-bionemo@users.noreply.github.com>

Author: svc-bionemo

bionemo-recipes/recipes/vllm_inference/esm2/install_vllm.sh

@@ -17,6 +17,6 @@ TORCH_CUDA_ARCH_LIST="$ARCH" MAX_JOBS="$MAX_JOBS" \
     uv pip install -r requirements/build.txt --system
 TORCH_CUDA_ARCH_LIST="$ARCH" MAX_JOBS="$MAX_JOBS" \
     uv pip install --no-build-isolation -e . --system
-pip install --upgrade "transformers[torch]"
+pip install --upgrade "transformers[torch]>=4.56.0,<5"
 
 echo "vLLM installed for arch $ARCH"