(PUP-9931) Use ruby threadlocal implementation

If we cycle out JRuby instances while using the Java ThreadLocal class,
we end up retaining references to the old JRuby instances. A simple
solution to this is to use the native ruby ThreadLocalVar
implementation, so that when a JRuby instance is GCd everything
stored on the ruby Thread class will be collected as well.

lib/puppet/context.rb

@@ -1,4 +1,4 @@
-require 'concurrent'
+require 'puppet/thread_local'
 
 # Puppet::Context is a system for tracking services and contextual information
 # that puppet needs to be able to run. Values are "bound" in a context when it is created
@@ -21,12 +21,12 @@ class DuplicateRollbackMarkError < Puppet::Error; end
 
   # @api private
   def initialize(initial_bindings)
-    @stack = Concurrent::ThreadLocalVar.new(EmptyStack.new.push(initial_bindings))
+    @stack = Puppet::ThreadLocal.new(EmptyStack.new.push(initial_bindings))
 
     # By initializing @rollbacks to nil and creating a hash lazily when #mark or
     # #rollback are called we ensure that the hashes are never shared between
     # threads and it's safe to mutate them
-    @rollbacks = Concurrent::ThreadLocalVar.new(nil)
+    @rollbacks = Puppet::ThreadLocal.new(nil)
   end
 
   # @api private
@@ -39,7 +39,7 @@ def push(overrides, description = '')
   #
   # @api private
   def unsafe_push_global(overrides, description = '')
-    @stack = Concurrent::ThreadLocalVar.new(
+    @stack = Puppet::ThreadLocal.new(
       @stack.value.push(overrides, description)
     )
   end

lib/puppet/thread_local.rb

@@ -0,0 +1,7 @@
+require 'concurrent'
+
+# We want to use the pure Ruby implementation even on JRuby. If we use the Java
+# implementation of ThreadLocal, we end up leaking references to JRuby instances
+# and preventing them from being garbage collected.
+class Puppet::ThreadLocal < Concurrent::RubyThreadLocalVar
+end