EP GB200 (4 GPUs/node) support

- configs.cuh: NUM_MAX_NVL_PEERS 8 -> 4
- internode.cu: introduce NvlPackT (uint64_t for 8 peers, uint32_t for 4)
  to handle packed-bool loads of is_token_in_rank; relax SourceMeta
  static_assert; replace 4 uint64_t-coupled sites
- buffer.hpp/buffer.cc: relax NUM_MAX_NVL_PEERS assert (4 || 8); read
  MSCCLPP_EP_LOCAL_WORLD_SIZE env to override rdma_rank/nvl_rank
  partitioning when local world size != NUM_MAX_NVL_PEERS
- CMakeLists.txt (ext/ep): rpath / install fix
- pyproject.toml: MSCCLPP_BUILD_EXT_EP=ON
- src/core/atomicadd_kernel.cu, kernels/buffer.cuh, kernels/utils.cuh:
  related EP fixes
- test_internode_multirank.py: NUM_MAX_NVL_PEERS=4, rank %% 4

test/python/ext/ep/test_internode_multirank.py

@@ -44,7 +44,7 @@ import torch.distributed as dist
 def init_dist():
     rank = int(os.environ["RANK"])
     world_size = int(os.environ["WORLD_SIZE"])
-    local_rank = int(os.environ.get("LOCAL_RANK", rank % 8))
+    local_rank = int(os.environ.get("LOCAL_RANK", rank % 4))
     torch.cuda.set_device(local_rank)
     dist.init_process_group(
         backend="nccl", world_size=world_size, rank=rank, device_id=torch.device(f"cuda:{local_rank}")
@@ -71,7 +71,7 @@ def main():
     rank, num_ranks, local_rank, group = init_dist()
     from mscclpp.ext import ep
 
-    NUM_MAX_NVL_PEERS = 8
+    NUM_MAX_NVL_PEERS = 4
     assert (
         num_ranks % NUM_MAX_NVL_PEERS == 0 and num_ranks > NUM_MAX_NVL_PEERS
     ), f"expected >1 node with 8 GPUs each, got num_ranks={num_ranks}"