This is a prerequisite of #1294.

This makes the alignment of promise a minimum of 8 (which was already the alignment on x86_64), which gives us 3 bits in the atomic state of a coroutine.

These 3 bits are used as follows (starting with LSB):

• Cancel bit. Set when a coroutine is canceled.
• Suspend bit. Set when a coroutine is suspended, and cleared when it is resumed.
• Awaited bit. Set when a coroutine is awaited.

I had to add more atomicrmw operations to make this all work. The atomic ops are now:

• suspend
  • atomicrmw OR itself (set the suspend bit, check if canceled)
• resume
  • atomicrmw AND target (clear the suspend bit, check if canceled)
• cancel
  • atomicrmw OR target (set the cancel bit)
• await
  • atomicrmw OR target (set the await bit and store its handle, check if canceled).
  • atomicrmw OR itself (set the suspend bit)
• return
  • atomicrmw OR itself (set the handle bits, check for await handle)
  • atomicrmw AND await handle (clear the suspend bit, check if canceled)

This fixes a few issues:

• closes improve coroutine cancel semantics #1261
• closes canceling a coroutine leads to defers not executed #1259
• closes support break in suspend blocks #803

#1163 is still a problem - an awaiting coroutine is indistinguishable from a suspended coroutine to zig.

With this, cancel becomes thread-safe, just like await already is.