Add all-gather + matmul ring tutorial

[yongweiy]

Summary

Adds a new tutorial under src/nki_samples/tutorials/allgather_matmul_ring/ demonstrating a fused all-gather + matmul along a ring of ranks, using nki.collectives.collective_permute_implicit (CPI) to overlap communication with compute.

Each ring step, a rank computes a local matmul against the LHS fragment currently in its ring buffer, then passes the fragment on to the next rank and receives the previous rank's fragment — the scheduler places the matmul of step i and the CPI of step i on disjoint engines so they run concurrently. The matmul is row-parallel (LHS row-sharded, RHS column-sharded). After RANK_N ring steps every rank has computed one (M_LOCAL, N_LOCAL) slot of the fully-gathered output for every source rank.

Follows the existing tutorial layout:

• allgather_matmul_ring_nki_kernels.py — NKI kernel (guarded with NKI_EXAMPLE_AGMM_RING_* markers). Module-level constants (RANK_N, M_LOCAL, N_LOCAL, K, K_TILE, N_TILE) keep the kernel signature small: allgather_matmul_ring(lhs_shard, rhs_shard, replica_group).
• allgather_matmul_ring_torch.py — PyTorch/XLA runner: xmp.spawn across ranks, builds deterministic LHS/RHS shards, validates each rank's output against a reference matmul computed on the host.

Sized for world_size=8, LNC=1 on a single trn2 device — the kernel has no LNC=2-specific code path, so LNC=1 is the canonical config (M_LOCAL=128 matches the Tensor Engine partition-dim limit of a single physical core, so each rank fully utilizes its core).

Test plan

• trn2, world_size=8, LNC=1 (single device, 8 cores):

  NEURON_CC_FLAGS="--lnc=1" \
  NEURON_LOGICAL_NC_CONFIG=1 \
  NEURONCORE_NUM_DEVICES=8 \
    python allgather_matmul_ring_torch.py
  

  All 8 ranks PASS with rel_err in 0.0006-0.0025 (threshold < 0.01).
• Profile inspection on NeuronExplorer confirms Tensor-engine matmul and CPI overlap on disjoint engines per ring step.

Other world_size × LNC combinations may fail if the replica group does not map to a valid CPI ring topology on the hardware; documented in the tutorial's module docstring.