Fix compiler output equivalence and add comprehensive tests

- Fix forloop.first/last/size lookups to try hash key before method call
- Fix tablerow cols parameter to use attributes['cols'] not @cols
- Fix tablerow output format to match interpreter (newlines, row boundaries)
- Fix capture compiler to access @body directly instead of iterating nodelist
- Add CompiledTemplate class to encapsulate code and external_tags
- Add external filter support via filter_handler
- Add debug mode warnings for external tag/filter calls
- Add Ruby 3.3 compatibility shim for peek_byte/scan_byte
- Add comprehensive unit tests for output equivalence
- Add benchmark comparing compiled vs interpreted rendering

All 30 test templates now produce identical output between compiled
Ruby and interpreted Liquid. Pre-compiled Ruby is 1.68x faster.

performance/ruby33_compat.rb

@@ -0,0 +1,26 @@
+# frozen_string_literal: true
+
+# Compatibility shim for Ruby 3.3
+# The Liquid library uses peek_byte and scan_byte which are only available in Ruby 3.4+
+
+require 'strscan'
+
+unless StringScanner.method_defined?(:peek_byte)
+  class StringScanner
+    def peek_byte
+      return nil if eos?
+      string.getbyte(pos)
+    end
+  end
+end
+
+unless StringScanner.method_defined?(:scan_byte)
+  class StringScanner
+    def scan_byte
+      return nil if eos?
+      byte = string.getbyte(pos)
+      self.pos += 1
+      byte
+    end
+  end
+end