Disable inlining for effect functions that manipulate stacks

[TheNumbat]

All of the effect continue functions extract a stack pointer from the given continuation and hide it in an immediate.
If we enter the GC between extracing the stack and calling resume, the stack will not be marked, leading to memory corruption.
This situation can only occur due to re-ordering allocations after inlining, so we can work around this issue by disabling inlining.

Additionally, Deep.try/match_with and Shallow.fiber must not enter the GC between alloc_stack and runstack for the same reason.

A better fix would be to change the fiber layout to not rely on storing the stack/last_fiber pointers as immediates, or at least use a single primitive to implement atomic continue and alloc_and_runstack operations.

[mshinwell]

I think this probably should have had comments to explain the reason for those attributes in the code.