Force the hash to include `librmm`

[jakirkham]

When inspecting recent libxgboost packages to find whether librmm was used to construct the hash, found the following results. Note that nccl is present, but librmm is missing.

```bash $ conda inspect hash-inputs libxgboost-1.7.6-cuda112h1c52686_0.conda {'libxgboost-1.7.6-cuda112h1c52686': {'recipe': {'__cuda': '__cuda', 'c_compiler': 'gcc', 'c_compiler_version': '10', 'channel_targets': 'conda-forge ' 'main', 'cuda_compiler': 'nvcc', 'cuda_compiler_version': '11.2', 'cxx_compiler': 'gxx', 'cxx_compiler_version': '10', 'nccl': '2', 'target_platform': 'linux-aarch64'}}} ```

While there are several ways to affect what is used in the package hash, this PR picks force_use_keys to ensure librmm is always included in the package hash. This didn't work as the librmm key was not found

Have added librmm to conda_build_config.yaml, which also should be used for hash inclusion. If that doesn't work, can try adding to force_use_keys with this change

cc @raydouglass

[jakirkham]

Ok looks like this is behaving as expected on CI:

{
  "libxgboost-1.7.6-rapidsai_h795055c": {
    "recipe": {
      "c_compiler": "gcc",
      "c_compiler_version": "10",
      "channel_targets": "rapidsai-nightly main",
      "cuda_compiler": "nvcc",
      "cuda_compiler_version": "11.2",
      "cxx_compiler": "gxx",
      "cxx_compiler_version": "10",
      "librmm": "23.10",
      "nccl": "2",
      "target_platform": "linux-64"
    }
  },
  ...
}

[jakirkham]

Can see the new libxgboost packages on rapidsai-nightly: