Fix: per-thread stack-pool for ExecutionContext::MultiThreaded

The current Fiber releases the dying fiber's stack to the pool before it
switches context to another fiber, which means there is a time window
during which a scheduler may pick the stack and start a new fiber while
another scheduler still hasn't switched context, which will corrupt the
stack.

src/multi_threaded.cr

@@ -15,7 +15,6 @@ abstract class ExecutionContext
     getter? idle : Bool = false
 
     protected getter global_queue : GlobalQueue = GlobalQueue.new
-    getter stack_pool : Fiber::StackPool = Fiber::StackPool.new
 
     # :nodoc:
     def self.default(size : Int32) : self
@@ -96,6 +95,11 @@ abstract class ExecutionContext
       unpark_idle_thread
     end
 
+    # TODO: there should be one stack pool per execution context (not per scheduler)
+    def stack_pool : Fiber::StackPool
+      raise "BUG: call #{self.class.name}#Scheduler#stack_pool instead of {self.class.name}#stack_pool"
+    end
+
     # TODO: there should be one event loop per execution context (not per scheduler)
     def event_loop : Crystal::EventLoop
       raise "BUG: call #{self.class.name}#Scheduler#event_loop instead of #{self.class.name}#event_loop"

src/multi_threaded/scheduler.cr

@@ -30,7 +30,9 @@ abstract class ExecutionContext
 
       # :nodoc:
       def stack_pool : Fiber::StackPool
-        @execution_context.stack_pool
+        # TODO: there should be one stack pool per execution context
+        # @execution_context.stack_pool
+        @stack_pool ||= Fiber::StackPool.new
       end
 
       # Unlike `ExecutionContext::MultiThreaded#enqueue` this method is only