Fix performance regression in do_write(s)

This causes significant performance issues when using large (>10meg) writes

Fix by adjusting the buffer write function to clear the buffer once, rather than
piece by piece, avoiding a case where a large write (in our case, around
70mbytes) will consume 100% of CPU. This takes a webrick GET request via SSL
from around 200kbyts/sec and consuming 100% of a core, to line speed on gigabit
ethernet and 6% cpu utlization.

lib/openssl/buffering.rb

@@ -348,13 +348,18 @@ def do_write(s)
     @wbuffer << s
     @wbuffer.force_encoding(Encoding::BINARY)
     @sync ||= false
-    if @sync or @wbuffer.size > BLOCK_SIZE
-      until @wbuffer.empty?
-        begin
-          nwrote = syswrite(@wbuffer)
-        rescue Errno::EAGAIN
-          retry
+    buffer_size = @wbuffer.size
+    if @sync or buffer_size > BLOCK_SIZE
+      nwrote = 0
+      begin
+        while nwrote < buffer_size do
+          begin
+            nwrote += syswrite(@wbuffer[nwrote, buffer_size - nwrote])
+          rescue Errno::EAGAIN
+            retry
+          end
         end
+      ensure
         @wbuffer[0, nwrote] = ""
       end
     end