Add UCCL test infrastructure and CI workflow

[adityakankariya]

Motivation

UCCL currently has no test infrastructure within TheRock's CI. The upstream UCCL project has its own CI workflow that runs multi-node tests via SSH, but TheRock needs a single-node, container-based test suite that follows the existing patterns established by the PyTorch external build (external-builds/pytorch/).

Technical Details

Adds a two-tier test suite for UCCL, mirroring the structure in external-builds/pytorch/:

New files in external-builds/uccl/:

• requirements-test.txt — test dependencies (pytest, numpy)
• smoke-tests/uccl_smoke_test.py — pytest smoke tests: ROCm availability, UCCL import verification, basic GPU tensor operations
• run_uccl_smoke_tests.py — smoke test runner with --amdgpu-family support and pytest passthrough
• run_uccl_tests.py — wraps upstream test_intranode.py via torchrun --standalone, exercising EP dispatch/combine kernels on a single node with auto-detected GPU count

New CI workflow:

• .github/workflows/test_uccl_wheels.yml — reusable workflow (workflow_dispatch + workflow_call) that runs smoke tests then intranode EP tests inside a container with GPU device passthrough, using the same container image and runner patterns as test_pytorch_wheels.yml

Updated:

• external-builds/uccl/README.md — added testing documentation

Test Plan

• uccl_repo.py checkout successfully clones UCCL sources with test_intranode.py at the expected path
• run_uccl_tests.py --dry-run --nproc-per-node 8 produces the correct torchrun command
• Workflow YAML passes syntax validation
• End-to-end GPU execution (requires CI runner with GPU access and a published UCCL wheel)

Test Result

Local validation passed on all non-GPU checks. The dry-run output confirms the correct torchrun invocation:

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.