The current implementation of parsing Cabal data is a big mess, which affects performance as described in #671 and #550. In this PR we simplify the implementation and avoid unnecessary reparsing.

As a result, the zero build time is reduced from 30 seconds to 5 seconds on my machine. The full build time seems to have been improved as well, since now all CI bots complete within the time limit. On the other hand, the speed up of the full build with `-j` on my machine does not appear to be significant.