Use a leaky singleton for CublasHandlePool. (pytorch#60987)

Summary:
Pull Request resolved: pytorch#60987

We were seeing deadlocks as follows during shutdown:

```
Thread 1 (LWP 2432101):
#0  0x00007efca470190b in __pause_nocancel () from /lib64/libc.so.6
#1  0x00007efca49de485 in __pthread_mutex_lock_full () from /lib64/libpthread.so.0
#2  0x00007ef91d4c42c6 in __cuda_CallJitEntryPoint () from /lib64/libnvidia-ptxjitcompiler.so.1
#3  0x00007efc651ac8f1 in ?? () from /lib64/libcuda.so
#4  0x00007efc651aee03 in ?? () from /lib64/libcuda.so
#5  0x00007efc64f76b84 in ?? () from /lib64/libcuda.so
#6  0x00007efc64f77f5d in ?? () from /lib64/libcuda.so
#7  0x00007efc64eac858 in ?? () from /lib64/libcuda.so
#8  0x00007efc64eacfbc in ?? () from /lib64/libcuda.so
#9  0x00007efc7810a924 in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#10 0x00007efc780fa2be in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#11 0x00007efc78111044 in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#12 0x00007efc7811580a in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#13 0x00007efc78115aa4 in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#14 0x00007efc781079ec in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#15 0x00007efc780e6a7a in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#16 0x00007efc7811cfa5 in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#17 0x00007efc777ea98c in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#18 0x00007efc777ebd80 in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#19 0x00007efc777ea2c9 in ?? () from /usr/local/cuda/lib64/libcublas.so.11
#20 0x00007efc778c2e2d in cublasDestroy_v2 () from /usr/local/cuda/lib64/libcublas.so.11
#21 0x00007efc51a3fb56 in std::_Sp_counted_ptr_inplace<at::cuda::(anonymous namespace)::DeviceThreadHandlePool<cublasContext*, &at::cuda::(anonymous namespace)::createCublasHandle, &at::cuda::(anonymous namespace)::destroyCublasHandle>, std::allocator<at::cuda::(anonymous namespace)::DeviceThreadHandlePool<cublasContext*, &at::cuda::(anonymous namespace)::createCublasHandle, &at::cuda::(anonymous namespace)::destroyCublasHandle> >, (__gnu_cxx::_Lock_policy)2>::_M_dispose() () from /data/users/pritam/pytorch/torch/lib/libtorch_cuda.so
#22 0x00007efc51a3fc5f in std::shared_ptr<at::cuda::(anonymous namespace)::DeviceThreadHandlePool<cublasContext*, &at::cuda::(anonymous namespace)::createCublasHandle, &at::cuda::(anonymous namespace)::destroyCublasHandle> >::~shared_ptr() () from /data/users/pritam/pytorch/torch/lib/libtorch_cuda.so
#23 0x00007efca4648b0c in __run_exit_handlers () from /lib64/libc.so.6
#24 0x00007efca4648c40 in exit () from /lib64/libc.so.6
#25 0x0000558c8852e5f9 in Py_Exit (sts=0) at /tmp/build/80754af9/python_1614362349910/work/Python/pylifecycle.c:2292
#26 0x0000558c8852e6a7 in handle_system_exit () at /tmp/build/80754af9/python_1614362349910/work/Python/pythonrun.c:636
#27 0x0000558c8852e742 in PyErr_PrintEx (set_sys_last_vars=<optimized out>, set_sys_last_vars=<optimized out>) at /tmp/build/80754af9/python_1614362349910/work/Python/pythonrun.c:646
#28 0x0000558c88540dd6 in PyRun_SimpleStringFlags (command=0x7efca4dc9050 "from multiprocessing.spawn import spawn_main; spawn_main(tracker_fd=9, pipe_handle=13)\n", flags=0x7ffe3a986110) at /tmp/build/80754af9/python_1614362349910/work/Python/pythonrun.c:457
#29 0x0000558c88540ead in pymain_run_command (cf=0x7ffe3a986110, command=<optimized out>) at /tmp/build/80754af9/python_1614362349910/work/Modules/main.c:420
#30 pymain_run_python (pymain=0x7ffe3a986220) at /tmp/build/80754af9/python_1614362349910/work/Modules/main.c:2907
#31 pymain_main (pymain=0x7ffe3a986220) at /tmp/build/80754af9/python_1614362349910/work/Modules/main.c:3460
#32 0x0000558c8854122c in _Py_UnixMain (argc=<optimized out>, argv=<optimized out>) at /tmp/build/80754af9/python_1614362349910/work/Modules/main.c:3495
#33 0x00007efca4632493 in __libc_start_main () from /lib64/libc.so.6
#34 0x0000558c884e5e90 in _start () at ../sysdeps/x86_64/elf/start.S:103
```

This was likely caused due to a static singleton that wasn't leaky. Following
the guidance in https://isocpp.org/wiki/faq/ctors#construct-on-first-use-v2 to
use a leaky singleton instead.
ghstack-source-id: 132847448

Test Plan: Verified locally.

Reviewed By: malfet

Differential Revision: D29468866

fbshipit-source-id: 89250594c5cd2643417b1da584c658b742dc5a5c

Author: pritamdamania

aten/src/ATen/cuda/CublasHandlePool.cpp

@@ -34,7 +34,13 @@ cublasHandle_t getCurrentCUDABlasHandle() {
   // See: https://github.com/pytorch/pytorch/pull/22405
   // This thread local unique_ptrs will be destroyed when the thread terminates,
   // releasing its reserved handles back to the pool.
-  static auto pool = std::make_shared<CuBlasPoolType>();
+
+  // Use a leaky singleton for the pool following standard practice around
+  // singletons: https://isocpp.org/wiki/faq/ctors#construct-on-first-use-v2
+  static auto pool = std::shared_ptr<CuBlasPoolType>(
+      new CuBlasPoolType(), [](CuBlasPoolType* p) {
+        // Leak the memory.
+      });
   thread_local std::unique_ptr<CuBlasPoolType::PoolWindow> myPoolWindow(
       pool->newPoolWindow());