Add NCCL Jacobi example

[nv-snidhan]

Description

Add an example that uses NCCL (via nccl4py) with Warp arrays for multi-GPU halo exchange: a distributed Jacobi solver for the Laplace equation, the NCCL counterpart to the mpi4py example from GH-475.

Context

Gives users a multi-GPU reference to compare against a single-GPU baseline, mirroring the NCCL variants in NVIDIA/multi-gpu-programming-models

[shi-eric]

This is addressed by 2aa75f8.

The change adds warp/examples/distributed/example_jacobi_nccl.py, a distributed Jacobi solver that:

• Decomposes the interior rows across MPI ranks and uses nccl4py point-to-point operations on Warp array rows for periodic nearest-neighbor halo exchange.
• Uses MPI for NCCL setup and global residual reduction.
• Captures each two-iteration step in a CUDA graph.
• Compares the distributed result and performance against a single-GPU reference, exiting with a nonzero status if the results differ.

Public API behavior:

• No public Warp API is added or changed. The example uses existing Warp array, stream, kernel, and CUDA graph APIs.

Test coverage:

• No automated test files were added or modified.
• The example includes an in-script correctness comparison against the single-GPU reference. The specialized MPI, NCCL, and multi-GPU requirements make the lack of a standard unit test reasonable, although hardware-backed smoke coverage would provide additional confidence.

The CHANGELOG now points users to the new example.