Shrink the inline storage in connect_awaitable (#1974)

* Shrink the inline storage in connect_awaitable

This diff adapts an idea originally due to @lewissbaker, originally
shared at https://godbolt.org/z/zGG9fsPrz.

Changes to Lewis's original include:
 * never invoke `coro.destroy()` since it's a no-op anyway
 * hard-code the integration with `connect_awaitable` since that's all
   we're using it for anyway
 * integrate with stdexec's support for various ways of making senders
   awaitable
 * MOAR `std::unreachable()`
 * support `unhandled_stopped()`

* Try a synthetic coroutine frame

This diff is inspired by the synthetic coroutine frame in @sgerbino's
fork of Capy, which I looked at here:
https://github.com/sgerbino/capy/blob/pr/beman-bench/bench/beman/awaitable_sender.hpp#L87-L92

The idea is to fake out a `coroutine_handle<__promise>` from a struct
like this:
```
struct __synthetic_coro_frame
{
  void (*__resume)(void*) noexcept;
  void (*__destroy_)(void*) noexcept;
};
```
and lie to the compiler with code like this:
```
// in __opstate<...>::__co_impl, which is a static member function
__std::coroutine_handle<__promise_t>::from_address(&__op.__synthetic_frame_);
```

With the right pointer math and a sprinkling of `reinterpret_cast`, this
code builds and passes all tests with Clang 22 and GCC 15 on Apple
silicon.

* Buy security at the cost of a pointer

The synthesized coroutine handle is passed to user code when it connects
an awaitable that suspends (it's received as the argument to
`await_suspend`), which means that user code can invoke `coro.destroy()`
even though we never do that in library code. The previous memory layout
stored `__opstate<...>::__started_` in the last byte of its synthesized
frame by artificially shortening the struct by a byte and marking the
`__started_` member with `[[no_unique_address]]`, which means that
`destroy` could point to a somewhat arbitrary address, depending on how
a `true` `bool` in the last byte is interpreted. This is potentially a
security hole so, to be safe, let's not squeeze so hard and just make
room for a function pointer and assign it the address of a no-op
function.

---------

Co-authored-by: Eric Niebler <eniebler@nvidia.com>

Author: ispeters

.github/workflows/test-windows.ps1

@@ -24,6 +24,7 @@ Invoke-NativeCommand cmake -B $BuildDirectory -G Ninja `
 	"-DCMAKE_MSVC_DEBUG_INFORMATION_FORMAT:STRING=Embedded" `
 	"-DSTDEXEC_ENABLE_ASIO:BOOL=TRUE" `
 	"-DSTDEXEC_ASIO_IMPLEMENTATION:STRING=boost" `
+	"-DCMAKE_VERBOSE_MAKEFILE:BOOL=ON" `
 	"-DSTDEXEC_BUILD_TESTS:BOOL=TRUE" .
 Invoke-NativeCommand cmake --build $BuildDirectory
-Invoke-NativeCommand ctest --test-dir $BuildDirectory
+Invoke-NativeCommand ctest --test-dir $BuildDirectory --output-on-failure --verbose --timeout 60