Fix an overflow with freq=4096 in the SSE4 decoder.

[jkbonfield]

Specifically, the SIMD decoder handles this case, but the tidyup code at the end for any remainder of < 32 bytes to go didn't.

Triggered by the follow example:

perl -e 'print "\0"x0xff0,"AAABAA\0"'|./tests/rans4x16pr -t -o5 -c 0x0201