[dnl] add NCCL/PT debug log for S413673

Test Plan:
Smoke test w/ NCCL ut cannot repro segfault w/ dynamic register + len 213942272
```
IFNAME=eth2 HOSTS="rtptest908.pci1,rtptest693.pci1" ENVS="NCCL_DEBUG=INFO;NCCL_DEBUG_SUBSYS=INIT,COLL,ALLOC" buck2 run fbcode//mode/opt fbsource//third-party/nccl-exp/v2.18.3-1/src/ctran/tests:ctran_dist_allgather
```
P1223852264

Differential Revision: D56659330

c10/cuda/CUDACachingAllocator.cpp

@@ -34,6 +34,8 @@
 #include <utility>
 #include <vector>
 
+#include <c10/util/Logging.h>
+
 TORCH_SDT_DEFINE_SEMAPHORE(malloc)
 TORCH_SDT_DEFINE_SEMAPHORE(free)
 
@@ -2796,6 +2798,12 @@ class DeviceCachingAllocator {
       cudaStream_t stream,
       c10::DeviceIndex device,
       std::shared_ptr<GatheredContext> context) {
+    if (action == TraceEntry::SEGMENT_ALLOC) {
+      LOG(INFO) << "CacheAllocator: SEGMENT_ALLOC addr=0x" << std::hex << addr
+                << std::dec << ", size=" << size
+                << ", trace_trackers_.size=" << trace_trackers_.size();
+    }
+
     if (!record_history && trace_trackers_.empty())
       return;
 

torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp

@@ -875,6 +875,7 @@ ProcessGroupNCCL::ProcessGroupNCCL(
   // lazyInitCUDA is called (and is a no-op if CUDA is already initialized).
   if (useTensorRegisterAllocatorHook_ && !allocatorHooksAttached) {
     at::globalContext().lazyInitCUDA();
+    LOG(INFO) << logPrefix() << "Registered cacheAllocatorRegisterHook";
     c10::cuda::CUDACachingAllocator::attachAllocatorTraceTracker(
         &cacheAllocatorRegisterHook);
     c10::cuda::CUDACachingAllocator::attachAllocatorTraceTracker(