All serialization is now down via CMark (previously, all `api` and `ucm` blocks in the output were still being
serialized directly, because of the interleaving of the input text and command/request output).

This causes a few changes to the existing transcript outputs:
- there is now always a blank line between Markdown block elements,
- leading blank lines in UCM blocks are gone,
- blank lines at the end of transcripts are gone (they still end with a final POSIX newline), and
- some `ucm` blocks now have 4-backtick fences (because they contain triple-backticks).

Transcript failures are now also handled as Markdown, rather than just being text appended to the document. This mostly
doesn’t change the serialization, except that the failure message is now fenced, since they often contain newlines and
indentation that is mangled otherwise.

This includes changes that get us closer to running transcript outputs as inputs
- all `ProcessedBlock`s (`api`, `ucm`, and `unison`) now handle info tags the same way – previously `ucm` didn’t allow
spaces between tags like `:hide :error`, and `api` didn’t support info tags at all.
- preserve info tags in the output – `ucm :error` in the input results in `ucm :error` in the output
- parser now supports `:added-by-unison` (only useful for parsing outputs as inputs)

- fix `:hide` on `ucm` and `api` blocks – previously it behaved the same as `:hide:all`, which prevents the outputs being used
as transcripts
- include `:added-by-ucm` to `ucm` blocks inserted after `unison` blocks
- ensure output lines have 2-space indents

This adds a couple transcripts testing the `:hide` behavior.

This provides the final pieces to make transcripts idempotent enough
that we can use transcript outputs as inputs.

This simplifies bug reporting, as now the GitHub issue can be written
_as_ a transcript, even if you want to include some of the output in the
bug report. No need to wrap the entire transcript in ` ```` markdown`.

The parser changes here also improve error messages in some existing cases.

There are some caveats:
- failing transcripts are not _yet_ idempotent. There are issues with how errors are reported (especially parse
failures) that make them not round-trip
- `:hide:all` almost always breaks idempotency. This is intentional – to just hide the output, use `:hide` instead.

These were previously separated because transcripts previously couldn’t contain code blocks with triple backticks, but that is fixed now.

This produces the first output files, which should remain unchanged on
subsequent runs.

Idempotent transcripts can get large, so using `:hide` more liberally is
probably a good idea. This tries that out on a few transcripts.

Checks that other transcripts in the source tree (currently just the GitHub
bug-report template) are valid.

This also fixes a bug where a transcript block like ` ``` ucm :hidec`
would be parsed like ` ``` ucm :hide c`, rather than complaining that
there is no `hidec` tag.

This currently fails because of a typo introduced earlier in this PR.

This is caught by a test introduced in the previous commit.

This also makes a couple minor changes re: running the script:

- removes the “belt and suspenders” `echo`ing that resulted in things being printed in triplicate
- added `gettext` to the Nix environment,  so `envsubst` is available
- changed the #!, to not get stuck with Bash 3.2 on macOS.

This avoids leading and trailing blanks, while ensuring they exist between commands & outputs.