OpenJDK6 and OpenJDK7 both have a bug in their xml encoder
(com.sun.org.apache.xml.internal.serializer.EncodingInfo), which can result in
a StackOverflowError.

The test case which tries to encode character 4096 to 321 easily reproduces the
issue. The issue appears to be this: When finding the encoding for (char)
4096, an EncodingImpl object is created which explicitly manages characters
Refer to https://bugs.openjdk.java.net/show_bug.cgi?id=100017

4096 to 4223. So far so good. But when the transformer tries to find the
encoding for (char) 4095, a new EncodingImpl delegate object is created which
manages the values 4095 to 4222. Effectively, this object only manages the
value 4095 (since the parent already manages 4096 to 4222). To find the value
for 4094, a new delegate is created. Do this a few more times, and you have one
stack frame for each value. Trying to encode characters from 0xffff to 0x0000
will result in thousands of stack frames.

The proposed patch makes sure that each delegate manages 128 values which dont
overlap with any other delegate. The patch brings down the maximum number of
delegates in the chain to (0xffff + 1)/RANGE = 512 (also the max number of
stack frames that can be used by EndcodingImpl).