⚡ Optimize the context layout and allocation strategy

[kammce]

Summary

Replace std::span<stack_word> m_stack with two raw pointers (m_stack_pointer, m_stack_end), saving one word in the context object and bringing hot fields into the first cache line. Capacity is stored in the sentinel slot at m_stack_end rather than in span metadata.

Changes

• Replace m_stack span with m_stack_pointer + m_stack_end pointers
• Store capacity in m_stack.back() and read it back via *m_stack_end
• Reorder members to pack m_state and m_sleep_time before the pointers, fitting more into the first cache line
• Fix inplace_context constructor to call initialize_stack_memory() in the body rather than the initializer list — member zero-init runs after the base constructor and would otherwise clobber the capacity sentinel written into m_stack.back()
• Update sync_wait tests to capture ctx.capacity() at construction instead of comparing against the raw template argument (capacity is now N - 1 due to the reserved sentinel slot)

Test Plan

• Pre-commit checks pass
• New/updated tests cover the changes
• Tested locally with conan create .