Cap CUDA-extension build parallelism (fix HF Spaces OOMKill, exit 137)

Set MAX_JOBS=1 and NVCC_THREADS=1 so mamba_ssm and selective_scan_cuda_oflex
compile sequentially. The HF Spaces build VM has ~16 GB RAM; parallel nvcc
workers (3-4 GB each) consistently OOMKill on the bench Space build.
Serial build adds ~10 minutes but completes reliably.

Dockerfile

@@ -12,6 +12,13 @@ ENV TORCH_CUDA_ARCH_LIST=8.9
 ENV CUDA_HOME=/usr/local/cuda
 ENV MPLCONFIGDIR=/tmp/matplotlib
 
+# Cap CUDA-extension build parallelism so the HF Spaces build VM (~16 GB
+# RAM) does not OOMKill during mamba_ssm / selective_scan_cuda_oflex
+# compilation. Each nvcc worker can use 3-4 GB; serial build is slower
+# (~25 min total) but reliable.
+ENV MAX_JOBS=1
+ENV NVCC_THREADS=1
+
 RUN apt-get update && apt-get install -y \
     python3.10 \
     python3.10-venv \