When marshalling, the current file index was not stored. Because of
this, a ';' was inserted multiple times for each file, even if the file
did not change.
When unmarshalling, the fileIdx determined by number of ';' was ignored.
Thus, if there were more than one file, all the positions would still
point to the first file.
This allows us to round-trip Body to JSON and back without any loss as
long as the expression source codes are always valid UTF-8, and we require
that during expression parsing anyway so that is a fine restriction.
The JSON encoding is a little noisy to read due to the extra annotations
required to be lossless (including source ranges) but still relatively
compact due to the base64-VLQ encoding of the source location information.
