Fix startup crash when a non-superuser tries to watch

[MaienM]

The issue was that the transaction failed but was not rolled back, causing all subsequent commands to fail. This just adds a rollback in the catch, and it adjusts the tests to expect this.

Fixes #368, replaces #369.

[calebmer]

It’s cool that this works 👍

I wouldn’t expect rollback to work outside of a transaction.

[benjie]

What voodoo is this?

Issuing ROLLBACK outside of a transaction block emits a warning and otherwise has no effect.
-- https://www.postgresql.org/docs/9.6/static/sql-rollback.html

[benjie]

Ah; the query is from:

https://github.com/calebmer/postgraphql/blob/6b3374baa974d8efd110970e4bd9431b8b3771f3/resources/watch-fixtures.sql

This query issues a begin; it seems that the error thrown prevents further statements being executed. This is unexpected to me... I would expect pgClient to execute all of the statements issued, even if errors occur, ultimately culminating in the commit; (which would perform an rollback due to the errors) and then return the results (i.e. throw the error).