chainlint: annotate original test definition rather than token stream

When chainlint detects problems in a test, such as a broken &&-chain, it prints out the test with "?!FOO?!" annotations inserted at each problem location. However, rather than annotating the original test definition, it instead dumps out a parsed token representation of the test. Since it lacks comments, indentations, here-doc bodies, and so forth, this tokenized representation can be difficult for the test author to digest and relate back to the original test definition. However, now that each parsed token carries positional information, the location of a detected problem can be pinpointed precisely in the original test definition. Therefore, take advantage of this information to annotate the test definition itself rather than annotating the parsed token stream, thus making it easier for a test author to relate a problem back to the source. Maintaining the positional meta-information associated with each detected problem requires a slight change in how the problems are managed internally. In particular, shell syntax such as: msg="total: $(cd data; wc -w *.txt) words" requires the lexical analyzer to recursively invoke the parser in order to detect problems within the $(...) expression inside the double-quoted string. In this case, the recursive parse context will detect the broken &&-chain between the `cd` and `wc` commands, returning the token stream: cd data ; ?!AMP?! wc -w *.txt However, the parent parse context will see everything inside the double-quotes as a single string token: "total: $(cd data ; ?!AMP?! wc -w *.txt) words" losing whatever positional information was attached to the ";" token where the problem was detected. One way to preserve the positional information of a detected problem in a recursive parse context within a string would be to attach the positional information to the annotation textually; for instance: "total: $(cd data ; ?!AMP:21:22?! wc -w *.txt) words" and then extract the positional information when annotating the original test definition. However, a cleaner and much simpler approach is to maintain the list of detected problems separately rather than embedding the problems as annotations directly in the parsed token stream. Not only does this ensure that positional information within recursive parse contexts is not lost, but it keeps the token stream free from non-token pollution, which may simplify implementation of validations added in the future since they won't have to handle non-token "?!FOO!?" items specially. Finally, the chainlint self-test "expect" files need a few mechanical adjustments now that the original test definitions are emitted rather than the parsed token stream. In particular, the following items missing from the historic parsed-token output are now preserved verbatim: * indentation (and whitespace, in general) * comments * here-doc bodies * here-doc tag quoting (i.e. "\EOF") * line-splices (i.e. "\" at the end of a line) Signed-off-by: Eric Sunshine <sunshine@sunshineco.com> Signed-off-by: Taylor Blau <me@ttaylorr.com>