[SymmetricMemory] introduce multicast support, multimem_all_reduce_ and multimem_one_shot_all_reduce (#133424)

### Summary
- Added multicast support to SymmetricMemory. If the cuda runtime and cuda driver have multicast support, SymmetricMemory associate all peer buffers with a multicast object and exposes the multicast virtual address.
- Implemented `multimem_all_reduce_` and `multimem_one_shot_all_reduce` based on the multicast support. The two variants shows different performance characteristic for different message size. We plan to use Inductor for collective algo selection (and required symmetric memory buffer allocation).

### Benchmark

8xH100 (non-standard version with HBM2e at 650W). NVSwitch V3 with NVLS support.

![image](https://github.com/user-attachments/assets/4998a16b-c2c0-4797-9dd0-1da2303df947)

![image](https://github.com/user-attachments/assets/278ad361-52cb-4864-82c6-bb67e8d0a3fe)

Differential Revision: [D61682507](https://our.internmc.facebook.com/intern/diff/D61682507)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133424
Approved by: https://github.com/yf225, https://github.com/weifengpy

BUILD.bazel

@@ -744,6 +744,7 @@ cc_library(
             "torch/csrc/cuda/nccl.cpp",
             "torch/csrc/distributed/c10d/intra_node_comm.cu",
             "torch/csrc/distributed/c10d/CUDASymmetricMemory.cu",
+            "torch/csrc/distributed/c10d/CUDASymmetricMemoryOps.cu",
             "torch/csrc/distributed/c10d/Utils.cu",
             "torch/csrc/distributed/c10d/quantization/quantization_gpu.cu",
         ],