Add NCCL system-provided targets

[cerisier]

Summary

• add public nccl_interface_library and nccl_system targets
• link system consumers through the unversioned NCCL linker interface without packaging the redistribution DSO
• add consumer smoke coverage and a configured-graph purity check across the CUDA matrix

Why

The system-provided target contract requires a semantic system wrapper whose closure uses interface imports instead of packaged shared-library imports. NCCL exposed shared and static variants but omitted that system family. Both the older 2.25.1 archive layout and the newer 2.31.2 layout provide lib/libnccl.so and no dedicated stub directory, so the unversioned shared-library interface is the correct link input.

Checks

• reproduced the missing public targets on merged main
• built nccl_interface_library and nccl_system across all six e2e CUDA platforms
• built nccl_system_smoke for the old and new NCCL archive families
• confirmed every configured system closure contains nccl_interface_library and no nccl_shared_library
• inspected the link action and ELF: libnccl.so is the link interface, NEEDED is libnccl.so.2, and no NCCL DSO is present in output runfiles
• passed root bazel build //..., NCCL tests, and the CUDA 13.3.1 consumer //:all build