Implement Syntax 0.8

[stasm]

Port of projectfluent/fluent.js#317.

[stasm]

@Pike, the changes in fluent/syntax are taken one-to-one from the JS implementation. There's one case where I needed to handle Python2 vs 3 explicitly: when unescaping escape sequences in StringLiterals, fluent.js uses String.fromCodePoint. In Python 2, we need unichr; in Python 3 that's simply chr.

I also realized that neither structure nor behavior tests have a test which would discover the need to use unichr vs. chr. So I added reference fixtures in a separate commit :) Two things to note:

• Just as in fluent.js, we need to skip two reference fixtures (leading_dots.ftl and variant_lists.ftl) due to how the tooling parsers reject the whole message if an attribute is broken.
• I added a processing function argument to BaseNode.to_json() so that I can remove the span property from the JSON object. I also use it to reset Junk annotations to [].

[Pike]

Nice, tested this locally a bit, gave it a quick run through c-l tests without surprises, too.

I did a side-by-side review of the parsers, only found a nit in the js side (issue filed).

I also compared the test fixtures, and they're nicely equal between fluent-syntax and python-fluent. There's one exception, and that's fixtures_structure/multiline_pattern, which I analyzed, and the error messages are different for bad message starts. fluent-syntax finds the missing =, python-fluent complains about value--attribute-less messages. That could also go away with something along the lines of #77.

[stasm]

Thanks! I'll merge this on Monday.

The fixtures differences are documented in https://github.com/projectfluent/python-fluent/blob/master/tests/syntax/README.md. I'd like to find a solution and remove them soon after this lands. Thanks again!