RJIT: Prefer STDERR over $stderr

When $stderr is redirected, you'll have no way to see why RJIT got
broken. This reference must always be the actual stderr.

lib/ruby_vm/rjit/compiler.rb

@@ -67,8 +67,8 @@ def compile(iseq, cfp)
       compile_block(asm, jit:, pc:)
       iseq.body.jit_entry = @cb.write(asm)
     rescue Exception => e
-      $stderr.puts "#{e.class}: #{e.message}"
-      $stderr.puts e.backtrace
+      STDERR.puts "#{e.class}: #{e.message}"
+      STDERR.puts e.backtrace
       exit 1
     end
 
@@ -110,7 +110,7 @@ def entry_stub_hit(entry_stub, cfp)
 
       return block.start_addr
     rescue Exception => e
-      $stderr.puts e.full_message
+      STDERR.puts e.full_message
       exit 1
     end
 
@@ -165,7 +165,7 @@ def branch_stub_hit(branch_stub, cfp, target0_p)
 
       return target.address
     rescue Exception => e
-      $stderr.puts e.full_message
+      STDERR.puts e.full_message
       exit 1
     end