bump NCCL floor to 2.18.1.1, relax PyTorch pin

[jameslamb]

Contributes to rapidsai/build-planning#102

Fixes #217

Notes for Reviewers

How I tested this

Temporarily added a CUDA 11.4.3 test job to CI here (the same specs as the failing nightly), by pointing at the branch from rapidsai/shared-workflows#246.

Observed the exact same failures with CUDA 11.4 reported in rapidsai/build-planning#102.

...
  + nccl                     2.10.3.1  hcad2f07_0                  rapidsai-nightly     125MB
...
./WHOLEGRAPH_CSR_WEIGHTED_SAMPLE_WITHOUT_REPLACEMENT_TEST: symbol lookup error: /opt/conda/envs/test/bin/gtests/libwholegraph/../../../lib/libwholegraph.so: undefined symbol: ncclCommSplit
sh -c exec "$0" ./WHOLEMEMORY_HANDLE_TEST 
./WHOLEMEMORY_HANDLE_TEST: symbol lookup error: /opt/conda/envs/test/bin/gtests/libwholegraph/../../../lib/libwholegraph.so: undefined symbol: ncclCommSplit
sh -c exec "$0" ./GRAPH_APPEND_UNIQUE_TEST 

(build link)

Pushed a commit adding a floor of nccl>=2.18.1.1. Saw all tests pass with CUDA 11.4 😁

...
  + nccl                     2.22.3.1  hee583db_1                  conda-forge          131MB
...
(various log messages showing all tests passed)

(build link)

[jameslamb]

docs build here have been failing like this:

LibMambaUnsatisfiableError: Encountered problems while solving:
  - nothing provides _python_rc needed by python-3.12.0rc3-rc3_hab00c5b_1_cpython

Could not solve for environment specs
The following packages are incompatible
├─ cpuonly is requested and can be installed;
├─ python 3.12**  is installable with the potential options
│  ├─ python [3.12.0|3.12.1|...|3.12.5] would require
│  │  └─ python_abi 3.12.* *_cp312, which can be installed;
│  └─ python 3.12.0rc3 would require
│     └─ _python_rc, which does not exist (perhaps a missing channel);
└─ pytorch 2.0.0**  is not installable because there are no viable options
   ├─ pytorch 2.0.0 would require
   │  └─ python_abi 3.8.* *_cp38, which conflicts with any installable versions previously reported;
   ├─ pytorch 2.0.0 would require
   │  └─ python >=3.10,<3.11.0a0 , which conflicts with any installable versions previously reported;
   ├─ pytorch 2.0.0 would require
   │  └─ cpuonly <0 , which conflicts with any installable versions previously reported;
   ├─ pytorch 2.0.0 would require
   │  └─ python >=3.8,<3.9.0a0 , which conflicts with any installable versions previously reported;
   ├─ pytorch 2.0.0 would require
   │  └─ python >=3.9,<3.10.0a0 , which conflicts with any installable versions previously reported;
   ├─ pytorch 2.0.0 would require
   │  └─ python_abi 3.10.* *_cp310, which conflicts with any installable versions previously reported;
   ├─ pytorch 2.0.0 would require
   │  └─ python_abi 3.11.* *_cp311, which conflicts with any installable versions previously reported;
   ├─ pytorch 2.0.0 would require
   │  └─ python_abi 3.9.* *_cp39, which conflicts with any installable versions previously reported;
   └─ pytorch 2.0.0 would require
      └─ __cuda, which is missing on the system.

(build link)

because:

• docker image rapidsai/ci-conda:latest was updated to Python 3.12 in update 'latest' tags to Python 3.12 images ci-imgs#188
• wholegraph is pinning pytorch=2.0.0 and there are not Python 3.12 packages for that version on conda-forge (https://anaconda.org/conda-forge/pytorch/files?version=2.0.0)

This PR proposes loosening the pin on pytorch here.

cugraph currently pins to pytorch>=2.0,<2.2.0a0 (code link), but might go to <2.4.0a0 in rapidsai/cugraph#4615 (comment).

[jakirkham]

Agree this seems sensible (especially for this release). In 24.12 we can look at relaxing this further

[vyasr]

I'll defer to a wholegraph dev on the pytorch change in case there are any reasons not to do that, but the nccl update looks fine.

[jameslamb]

Thanks!

@linhu-nv could you please review here?

[linhu-nv]

This seems good to me. Thanks!

[jameslamb]

/merge