support mp on xpu (#49815)

1 update xccl lib
2 when using comm_ctx, the allocator should be set manually.

cmake/external/xpu.cmake

@@ -16,7 +16,7 @@ else()
 endif()
 
 set(XPU_XCCL_BASE_URL
-    "https://klx-sdk-release-public.su.bcebos.com/xccl/release/1.0.0")
+    "https://klx-sdk-release-public.su.bcebos.com/xccl/release/1.0.6")
 
 if(WITH_AARCH64)
   set(XPU_XRE_DIR_NAME "xre-kylin_aarch64")

paddle/fluid/distributed/collective/process_group_bkcl.cc

@@ -128,6 +128,11 @@ void ProcessGroupBKCL::CreateBKCLEnvCache(const Place& place,
       platform::DeviceContextPool::Instance().Get(place));
   // must use XPUDeviceContext here to make sure XPUContext::Init() is called
   auto comm_ctx = std::make_unique<XPUDeviceContext>(place);
+  // set allocator
+  comm_ctx->SetAllocator(memory::allocation::AllocatorFacade::Instance()
+                             .GetAllocator(place)
+                             .get());
+
   BKCLContext_t bkcl_comm;
   BKCLCHECK(bkcl_init_rank(&bkcl_comm, GetRank(), GetSize(), &bkcl_id));
   comm_ctx->SetBkclContext(bkcl_comm);