Add option to respect default-extensions from .cabal files

[amesgen]

Closes #517
Closes #519

TODO:

• Allow to explicitly set the path to the cabal file and to the source file (for input via stdin, for example for editor integrations)
• Respect language (in advance of the GHC2021 language in GHC 9.2)
• Test if this change significantly decreases performance. When ormolu is called with multiple files, we could cache the parsed .cabal files. EDIT see this comment
• Add tests (could be further extended)
• What do we want to do when the same file is part of different components? Right now, some arbitrary component is selected (via Map.union).
• Make this new functionality discoverable, e.g. in the README.

This PR already revealed one issue, #758.

[amesgen]

Concerning performance: This change is opt-in, so nothing will change in the default configuration. When -e is enabled, a small cost, dependent on the depth of the directory tree and the complexity of the .cabal file, has to be payed. When formatting ormolu, this means (using hyperfine and fd):

 $ hyperfine 'fd -e hs . src -X ormolu --mode check'
Benchmark #1: fd -e hs . src -X ormolu --mode check
  Time (mean ± σ):      1.148 s ±  0.031 s    [User: 1.119 s, System: 0.029 s]
  Range (min … max):    1.099 s …  1.198 s    10 runs

 $ hyperfine 'fd -e hs . src -X ormolu -e --mode check'
Benchmark #1: fd -e hs . src -X ormolu -e --mode check
  Time (mean ± σ):      1.291 s ±  0.032 s    [User: 1.250 s, System: 0.042 s]
  Range (min … max):    1.233 s …  1.323 s    10 runs

Considering that low hanging fruits like parallel formatting would make both cases much faster

 $ hyperfine 'fd -e hs . src -x ormolu --mode check'
Benchmark #1: fd -e hs . src -x ormolu --mode check
  Time (mean ± σ):     712.2 ms ±  40.7 ms    [User: 5.590 s, System: 0.981 s]
  Range (min … max):   662.1 ms … 790.8 ms    10 runs

 $ hyperfine 'fd -e hs . src -x ormolu -e --mode check'
Benchmark #1: fd -e hs . src -x ormolu -e --mode check
  Time (mean ± σ):     837.7 ms ±  20.6 ms    [User: 7.159 s, System: 1.104 s]
  Range (min … max):   810.9 ms … 872.6 ms    10 runs

I think that we don't have to worry too much about a performance regression in this PR.

[amesgen]

Updated (01-10-2021): See vyorkin/ormolu.el#14 for a way to use this feature with ormolu.el.