fix(stdexec): initialize stop-when's once-flag before its callbacks

[alwaysprince05]

Summary

Closes #2326. This reorders the once-flag __called_ in __stop_when_::__fused_token<T1, T2>::callback_type so that its lifetime starts before the two inner stop callbacks that use it.

Note: PR #2327 fixes the same bug the same way and is approved but unmerged; this PR was prepared independently and introduces a regression test that #2327 does not include.

Problem

callback_type declared its once-flag after the two inner callbacks:

_Fn                 __fn_;
__cb1_t             __cb1_;
__cb2_t             __cb2_;
__std::atomic<bool> __called_{false};   // initialized last

If the first token is already stopped, __cb1_'s constructor invokes the function immediately. That function does __called_.exchange(true) on an object whose lifetime has not begun, and then __called_'s own initializer stores false over it. Depending on the indeterminate value observed:

• 0: the function runs at registration, the flag is reset, and the function runs a second time when the second token is stopped;
• non-zero: the function is skipped at registration, and runs only if the second token is stopped later — it is lost otherwise.

Both violate [exec.stop.when]: the function must run when either token is stopped, and at most once. Regression from #2234.

Fix

__called_ is declared before __cb1_/__cb2_, so any immediate invocation during construction operates on a live, initialized flag:

_Fn                 __fn_;
__std::atomic<bool> __called_{false};
__cb1_t             __cb1_;
__cb2_t             __cb2_;

Verification

• The issue's reproducer prints calls at registration: 1 (expected 1), in total: 2 (expected 1) before this change, and the correct 1 and 1 after it.
• A regression test in test/stdexec/algos/adaptors/test_stop_when.cpp ("callbacks registered with an already-stopped fused token run exactly once") fails on main and passes with this change.
• test.stdexec (629 cases) green in Release and Debug, and with STDEXEC_ENABLE_EXTRA_TYPE_CHECKING on.
• clang-format 21 clean on both changed files.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[alwaysprince05]

/ok to test 9a04fc9

[ericniebler]

/ok to test 9a04fc9

[ericniebler]

this is a dupe of #2327, but i'm keeping it for the test.