Passing parent stack frames to spawn-helper functions

[neboat]

This PR drafts runtime-system changes to pass a pointer to the __cilkrts_stack_frame in a parent spawning function, along with other compile-time information, to a spawn helper. This change allows the spawn helper to access the parent stack frame more efficiently and enables a variety of optimizations to reduce spawn overhead.

• Spawn helpers can avoid looking up the parent stack frame from worker-local storage, which can be costly to access.
• Spawn-helper stack frames need not explicitly store a pointer to the parent stack frame.
• By receiving the compile-time information about whether a spawn helper is itself a spawner, the spawn-helper routines can avoid unnecessary initialization and deinitialization work, such as adding themselves to the cactus stack of __cilkrts_stack_frames. In this case, the compiler is able to optimize away the spawn-helper's stack frame.

This PR requires the compiler changes in OpenCilk/opencilk-project#252.

[VoxSciurorum]

This part looks good. Do you have any performance measurements?

[neboat]

I took some performance measurements on a couple AWS servers, using the Cilk-5 benchmarks and some older Intel codes. This change does improve performance a little bit on those benchmarks, though many of them may have been tuned already to minimize spawn overhead. On benchmarks that show more spawn overhead (cholesky and a coarsened fib), the improvement seems to be around 1–2% on Intel Cascade Lake and 2–4% on an M2.