Contributes to rapidsai/build-planning#54.

The `libraft-headers` and `libraft-headers-only` conda packages are bundling `rmm`'s headers. I believe that's because the `librmm` conda package isn't available in the `libraft*` conda build environment, and as a result it's getting `rmm` via CPM (thanks to `rapids-cmake`).

As a result, this project and any that depend on it are seeing warnings like the following in conda builds where `conda`'s `path_conflict` setting is set to `warn` or `prevent` (like #2245):

```text
This transaction has incompatible packages due to a shared path.
  packages: rapidsai-nightly/linux-64::librmm-24.04.00a38-cuda11_240326_ga98931b9_38, rapidsai-nightly/linux-64::libraft-headers-only-24.04.00a93-cuda11_240326_g9637b3c2_93
  path: 'include/rmm/mr/device/arena_memory_resource.hpp'
```

To fix that, this proposes the following changes:

* make `librmm` a `host` dependency of the following conda packages: `libraft-headers-only`, `libraft-headers`

### Benefits of this change

* slightly smaller `libraft-headers` and `libraft-headers-only` conda packages
* reduced risk of runtime and installation issues caused by file clobbering

## Notes for reviewers

### History of changes to the `librmm` dependency for `libraft-headers`:

* both `run` and `host`: #508
* both `run` and `host`, but ignoring its `run_exports`: #1264
* just `run`, but ignoring its `run_exports`: #2102

In particular, #2102 referred to the `host` dependency on `librmm` as "extraneous" but from a packaging perspective, I don't think it is. `librmm` being in `host` means it'll be present in the build environment, which means its headers will be *found* instead of *downloaded*, and therefore not packaging into the `libraft*` conda packages.

### How I tested this

Built all the `raft` conda packages locally from `branch-24.06` and confirmed that they contain `rmm` headers. Then again from this branch and confirmed they were gone.

```shell
docker run \
    --rm \
    --env-file "${PWD}/aws-creds.env" \
    -v $(pwd):/opt/work \
    -w /opt/work \
    -it rapidsai/ci-conda:cuda12.2.2-ubuntu22.04-py3.10-amd64 \
    bash

CI="true" \
  ci/build_cpp.sh

# On 'branch-24.06', this showed the rmm headers being packaged.
# On this branch, they're omitted.
tar --bzip2 -tf \
  /tmp/conda-bld-output/linux-64/libraft-headers-only-24.06.00a50-cuda12_240430_g1e0e2283_50.tar.bz2 \
| grep 'include/rmm' \
| wc -l
```

Also checked the CI logs from `conda-cpp-build` jobs here. On other recent PRs, I see CPM downloading `rmm` ...

```text
-- CPM: Adding package [email protected] (branch-24.06)
```

... and all the `rmm` headers getting installed as part of the `libraft-headers` package

```text
-- Installing: /opt/conda/conda-bld/_h_env_placehold_placehold_..._placeho/include/rmm/cuda_stream.hpp
```

([build link](https://github.com/rapidsai/raft/actions/runs/8904352932))

Here, I see `librmm` coming through via the conda package requirements ...

```text
The following NEW packages will be INSTALLED:
    ...
    librmm:                      24.06.00a17-cuda12_240430_g26fa9ecb_17 rapidsai-nightly
```

... and being used instead of downloads via CPM ...

```text
-- CPM: Using local package [email protected]
```

... and no installation of the `rmm` headers as part of building any `libraft` packages.

([build link](https://github.com/rapidsai/raft/actions/runs/8910675575/job/24470450187?pr=2284))

Authors:
  - James Lamb (https://github.com/jameslamb)

Approvers:
  - Bradley Dice (https://github.com/bdice)
  - Ray Douglass (https://github.com/raydouglass)

URL: #2284