Add stripped unit tests for decoder

That is, test cases where there's no newline or other whitespace at
the beginning or end of input. This was implemented by adding a
--strip option to split-testfile to strip the input file after writing
it.

The actual test JSON texts are the same as testdata/invalid and
testdata/valid. The expected output of the invalid cases had to be
adjusted a bit: because there's no newline at the end, some of the
line numbers needed to be changed.