TL/CUDA: raise MAX_NVLS_PEERS to 576 and guard nvls_init

GB200/NVL systems support up to 72 GPUs per NVSwitch domain with up
to 8 NVLink partitions per domain, yielding a theoretical maximum of
576 participants in a single NVLS multicast group.  Raise
UCC_TL_CUDA_MAX_NVLS_PEERS from 144 to 72*8=576 to accommodate these
configurations.

Add a guard in nvls_init STATE_INIT that rejects teams exceeding this
limit with UCC_ERR_NOT_SUPPORTED and a clear warning, preventing
out-of-bounds accesses in the NVLS allgather buffer.

Author: ikryukov

src/components/tl/cuda/tl_cuda.h

@@ -32,7 +32,8 @@
 #define UCC_TL_CUDA_MAX_RING_CHUNKS 8
 
 #ifdef HAVE_NVLS
-#define UCC_TL_CUDA_MAX_NVLS_PEERS 144
+/* 72 GPUs per NVSwitch domain * 8 partitions = 576 max NVLS peers */
+#define UCC_TL_CUDA_MAX_NVLS_PEERS (72 * 8)
 #define UCC_TL_CUDA_MAX_NVLS_SM_COUNT 32
 #define UCC_TL_CUDA_MAX_NVLS_THREADS 1024
 

src/components/tl/cuda/tl_cuda_nvls.c

@@ -329,6 +329,14 @@ ucc_status_t ucc_tl_cuda_nvls_init(
             return status;
         }
 
+        if (UCC_TL_TEAM_SIZE(team) > UCC_TL_CUDA_MAX_NVLS_PEERS) {
+            tl_warn(lib,
+                    "NVLS: team size %u exceeds maximum supported peers %d; "
+                    "disabling NVLS for this team",
+                    UCC_TL_TEAM_SIZE(team), UCC_TL_CUDA_MAX_NVLS_PEERS);
+            return UCC_ERR_NOT_SUPPORTED;
+        }
+
         if (nvls->is_multinode) {
             ucc_team_t *ucc_team = UCC_TL_CORE_TEAM(team);
             ucc_rank_t  min_ppn, max_ppn;