chore(nccl-tests): update NCCL and EFA stack

[KeitaW]

Purpose

Update the micro-benchmarks/nccl-tests container to the current EFA-tested NCCL stack and remove floating dependency references from this benchmark subtree.

This does not duplicate an existing PR. I searched open PRs for nccl-tests 2.31.2 and aws-ofi-nccl 1.21.1; no open PR changes this subtree. PR #1240 mentions aws-ofi-nccl 1.21.1 in a separate DeepEP example.

Changes

• Update NCCL to v2.31.2-1 and EFA installer to 1.50.0, which bundles aws-ofi-nccl v1.21.1 and libfabric 2.6.0amzn1.0.
• Update directly coupled dependencies to CUDA 13.1.2, GDRCopy v2.6, and nccl-tests v2.20.0.
• Pin the CUDA base image by its versioned tag and pin Python packages used in the image and topology test environment.
• Remove the floating :latest publication from the buildspec. The GB200 manifest remains unchanged and is outside this PR.
• Update both READMEs with release links, the exact image tag, and copy-pasteable build instructions.

Test Plan

Environment:

• AWS Service: local Docker build only; no AWS runtime was used
• Instance type: CPU build host, no GPU or EFA device
• Number of nodes: 1 build host

Test commands:

docker buildx build --platform linux/amd64 --load -t nccl-tests:validation -f nccl-tests.Dockerfile .
docker run --rm --entrypoint /bin/bash nccl-tests:validation -lc 'set -e; test "$(git -C /opt/nccl describe --tags --exact-match)" = v2.31.2-1; test "$(git -C /opt/nccl-tests describe --tags --exact-match)" = v2.20.0; strings /opt/amazon/ofi-nccl/lib/libnccl-net-ofi.so | grep -F "aws-ofi-nccl 1.21.1"; test -x /opt/nccl-tests/build/all_reduce_perf; test -x /opt/nccl-tests/build/alltoall_perf'
python3 -m venv /tmp/nccl-tests-req-venv
/tmp/nccl-tests-req-venv/bin/pip install -r micro-benchmarks/nccl-tests/slurm/topology-aware-nccl-tests/test_requirements.txt
/tmp/nccl-tests-req-venv/bin/pytest micro-benchmarks/nccl-tests/slurm/topology-aware-nccl-tests/test_hostfile_topologify.py -q
bash -n micro-benchmarks/nccl-tests/slurm/*.sbatch micro-benchmarks/nccl-tests/slurm/topology-aware-nccl-tests/*.sh micro-benchmarks/nccl-tests/slurm/topology-aware-nccl-tests/*.sbatch
python3 -m py_compile micro-benchmarks/nccl-tests/*.py micro-benchmarks/nccl-tests/slurm/topology-aware-nccl-tests/*.py
git diff --check origin/main...HEAD

Test Results

• linux/amd64 Docker build: PASS. Built image sha256:279c86a70d0bfca7d3ecf9a5d517814a724ebf944457abf72c4c738250981258.
• Image static verification: PASS. The built image reports NCCL v2.31.2-1, nccl-tests v2.20.0, and NET/OFI Initializing aws-ofi-nccl 1.21.1; all_reduce_perf and alltoall_perf exist and are executable.
• Pinned-requirement unit tests: PASS, 3 tests passed in 0.23 seconds.
• YAML parsing: PASS, 3 YAML files parsed.
• Shell syntax, Python byte-compilation, and git diff --check: PASS.
• Markdown lint reports 10 pre-existing errors in the 2 README files; this change introduces no new lint category or location outside those existing README defects.
• GPU collective execution, EFA transport selection, multi-node performance, and linux/arm64 image build were not run. The GB200 manifest is unchanged and outside this PR. This PR does not claim hardware runtime validation.

Directory Structure

No new directory is added. The update is limited to micro-benchmarks/nccl-tests.

Checklist

• I have read the contributing guidelines.
• I am working against the latest main branch.
• I have searched existing open and recently merged PRs to confirm this is not a duplicate.
• The contribution is self-contained with documentation and scripts.
• External dependencies are pinned to a specific version or tag (no latest).
• A README is included or updated with prerequisites, instructions, and known issues.
• New test cases follow the expected directory structure. No new test case was added.
• Test Results above contain e2e evidence from a real AWS run (or this PR is marked docs-only). Hardware e2e validation was not available; the exact build and static verification performed are reported above.