build: install fast-hadamard-transform with no build isolation

megatron-core[dev] (pulled in via the bumped Megatron-Bridge rev) now depends
on fast-hadamard-transform, a Dao-AILab CUDA extension that imports torch in
setup.py. Under build isolation uv builds it in a torch-less env, so the
container build fails with "ModuleNotFoundError: No module named 'torch'".

The previous extra-build-dependencies fix is silently ignored by the
CI-pinned uv 0.7.2 (unknown field warning). Add the package to
no-build-isolation-package instead, matching the other torch-at-build-time
packages (transformer-engine, mamba-ssm, causal-conv1d, nv-grouped-gemm);
torch is available from the PyTorch base image via --system-site-packages.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>

Author: ko3n1g

pyproject.toml

@@ -123,6 +123,7 @@ no-build-isolation-package = [
     "mamba-ssm",
     "causal-conv1d",
     "nv-grouped-gemm",
+    "fast-hadamard-transform",
 ]
 
 # Always apply the build group since dependencies like TE/mcore/nemo-run require build dependencies