add static_assert that unique_ptr<T> is constructible from not_null<unique_ptr<T>>

Summary:
Adding this as a test for the new compiler. With this upstream PR: llvm/llvm-project#136203

This test starts to fail with the following error:
```
fbcode/folly/memory/test/not_null_test.cpp:465:24: error: call to constructor of 'std::unique_ptr<int>' is ambiguous
  465 |   std::unique_ptr<int> i{std::move(nnu)};
      |                        ^~~~~~~~~~~~~~~~~
fbcode/third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/unique_ptr.h:327:7: note: candidate constructor
  327 |       unique_ptr(unique_ptr&&) = default;
      |       ^
fbcode/third-party-buck/platform010/build/libgcc/include/c++/trunk/bits/unique_ptr.h:468:7: note: candidate constructor has been explicitly deleted
  468 |       unique_ptr(const unique_ptr&) = delete;
      |       ^
fbcode/folly/memory/test/not_null_test.cpp:467:7: error: static assertion failed due to requirement 'std::is_constructible_v<std::unique_ptr<int, std::default_delete<int>>, folly::not_null<std::unique_ptr<int, std::default_delete<int>>>>'
  467 |       std::is_constructible_v<
      |       ^~~~~~~~~~~~~~~~~~~~~~~~
  468 |           std::unique_ptr<int>,
      |           ~~~~~~~~~~~~~~~~~~~~~
  469 |           not_null<std::unique_ptr<int>>>);
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2 errors generated.
```
Note that when you construct a unique_ptr, its deleted copy constructor still participates in overload resolution. So it's possible for a operator const PtrT&() const& noexcept to be selected. Meanwhile, its operator PtrT&&() && noexcept; is selected to convert not_null to the unique_ptr&& for the move constructor. Hence the ambiguity.

Why is this only a problem now? See D74599633. The existence of the conversion operator template was a better match to create a unique_ptr from a not_null previously. However, the SFINAE conditions are circular and caused weird compiler behaviours. This upstream PR now stops looking at templates when there's a "perfect match", which are the nontemplated `operator PtrT&&` and `operator const PtrT&` for not_null. It's possible that the PR has hidden some bugs, but the behaviour is unspecified by the standard anyway. The PR implements the change to make it align with GCC behaviour.

Reviewed By: snarkmaster, Gownta

Differential Revision: D74602687

fbshipit-source-id: ad6c49e6fa57c6340d6646734e9b6120e309353c

Author: yuxuanchen1997

third-party/folly/src/folly/memory/test/not_null_test.cpp

@@ -461,6 +461,13 @@ TEST_F(NotNullHelperTest, output) {
 }
 
 TEST_F(NotNullHelperTest, casting) {
+  auto nnu = make_not_null_unique<int>(42);
+  std::unique_ptr<int> i{std::move(nnu)};
+  static_assert(
+      std::is_constructible_v<
+          std::unique_ptr<int>,
+          not_null<std::unique_ptr<int>>>);
+
   not_null_shared_ptr<Derived> nnd(new Derived());
 
   auto s = static_pointer_cast<Base>(nnd);