Use nccl's run_export to add it to run

As `libxgboost` dynamically links to `nccl`, it must have the library
available to load it (even in the CPU case). So move `nccl` to a hard
dependency (instead of a soft one).

recipe/meta.yaml

@@ -49,7 +49,6 @@ outputs:
       string: rapidsai_h{{ PKG_HASH }}_{{ PKG_BUILDNUM }}
       ignore_run_exports_from:
         - {{ compiler('cuda') }}  # [(cuda_compiler_version or "").startswith("11")]
-        - nccl                    # [linux and cuda_compiler != "None"]
     requirements:
       build:
         - {{ compiler('c') }}
@@ -69,7 +68,6 @@ outputs:
         - cuda-version >=11.2,<12  # [(cuda_compiler_version or "").startswith("11")]
       run_constrained:
         - librmm ={{ rapids_version }}               # [linux and cuda_compiler != "None"]
-        - {{ pin_compatible("nccl", max_pin="x") }}  # [linux and cuda_compiler != "None"]
 
   - name: py-xgboost
     script: install-py-xgboost.sh  # [not win]