All of the standard tokenizer errors are implemented with their given
error code strings.

Each transition of the tokenizer state machine that emit one or more
errors is tested to emit exactly those errors and no others. Those
transitions that do not emit an error are tested to ensure that they do
not.

The `GumboError` is opaque to callers and the `GumboStringBuffer` is not
exposed. The error codes of the defined errors as well as the two
encoding errors and a generic parser error are available via
`gumbo_error_code`.

In Nokogiri, the error code is exposed via
`Nokogiri::XML::SyntaxError#str1`.

Replace

```
if (foo) {
  return x();
} else if (bar) {
  return y();
} else {
  return z();
}
```

with

```
if (foo) {
  return x();
}
if (bar) {
  return y();
}
return z();
```

The parser now implements the updated tree-construction stage including
producing the same number of errors specified in the html5lib-tests
tree-construction tests (at least it will after the tests are fixed as
per
https://github.com/html5lib/html5lib-tests/compare/master...stevecheckoway:all-error-fixes?expand=1).