Add a missing initialization to LZDecoder's constructor. The same
initialization is done in LZDecoder.reset(). LZMA2InputStream
always calls it before decoding anything, thus this bug didn't break
LZMA2 decoding. However, LZMAInputStream has no reason to call
LZDecoder.reset(), and thus the missing initialization in the
constructor meant that ArrayCache didn't work with LZMAInputStream.

Fixes: c5569e4 ("Add ArrayCache support to LZMA and LZMA2 coders, part 1.")
Fixes: #23

The extra size constant from UnalignedLongLEMatchLengthFinder didn't take
effect as intended. This could result in ArrayIndexOutOfBoundsException.
Frequent calls to flush() increased the likelyhood of hitting the bug
but it could be triggered without flushing too:

    ant clean jar
    head -c12853521 /dev/zero \
        | java -ea -jar build/jar/XZEncDemo.jar > /dev/null

The bug is present in versions 1.10 and 1.11. If one cannot upgrade,
one should set the property org.tukaani.xz.MatchLengthFinder=Basic
which disables UnalignedLongLEMatchLengthFinder. Example:

    head -c12853521 /dev/zero \
        | java -Dorg.tukaani.xz.MatchLengthFinder=Basic \
               -jar build/jar/XZEncDemo.jar > /dev/null

Fixes: 9fd7bc0 ("Add MatchLength.EXTRA_SIZE and use it in LZEncoder")
Fixes: #24

It's simpler code.