This stashes the HTML version of every entry in memory and uses
Mojo::DOM to extract headers from the markup for use as titles.
Titles are displayed in $self->{page_navigation}, now available
inside templates as ${page_navigation}.

In order to keep Mojo::DOM from choking on other input, it uses the open
pragma to open everything as UTF-8 by default, which also eliminates a
whole class of character encoding bugs and removes some fiddling with
Encode::decode() from feed_print().

This is all obviously a more memory-intensive, but caching the markup
turns out to have the side effect of making it much faster to render
even a large site, probably as much as anything because the HTML in
question is only getting generated once per entry instead of
(potentially) 2-3 times.

This commit isn't very atomic.  In the process of roughing it out and
testing it, I made a small pile of minor but potentially breaking
changes:

  - Removed entry_map from settings and hardcoded handling of various
    types of entry as some if-statements instead.

  - Removed embedded_perl flag in settings - was always turned on in
    practice, and wasn't very coherent since templating would have
    broken without it.

  - bin/wrt-display - now handles the "feed" alias correctly

  - EntryStore: now supports retrieving values for properties with
    prop_value() - this isn't currently used, but it seems like a
    reasonable extension of the property idea.

  - Added `wrt ls --with-titles`.

  - Added dependency versions to Build.PL.

  - Refactored Markup's line_parse() a little.

  - Refactored some tests to give cleaner / more useful output.

  - Renamed default template file to "default".

This commit is something of a hairball, the result of evenings-and-weekends
hacking building up a set of changes that got out of hand in parallel.
If I had the energy to spare, I would break it apart into
semantically-related changes, but I don't - and I suppose all this crap
being rolled together is at least reflective of how the code was
written.

These changes are really half-finished, at best.  Eventual goals:

  - App::WRT shouldn't directly touch the filesystem
  - App::WRT::EntryStore should model the entry archive completely
  - App::WRT::Renderer should say what to write to the publication
    directory
  - This one's a maybe: Filesystem interaction should pass through
    App::WRT::FileIO or something like it so that EntryStore and Renderer
    can be more usefully tested, with mocked writes (maybe)

I do think this represents an inflection point in the long, silly life of
this program:  It includes a handful of new tests, and a number of the
code changes were in turn easy to make because the test suite begins to
model the code in a useful way.  It's less and less necessary to run wrt
against the p1k3.com archives to be sure that I haven't trashed something.

Breaking changes to note:

  - Will no longer render HTML for nonexistent entries
  - Months and years which are flatfiles or contain an index are handled
    differently, albeit less brokenly
  - EntryStore includes index files in its overall list of entries
    (this seems to break less than I thought), which trickles out to
    bin/wrt-ls

Overall changes herein:

  - App::WRT::Date
    - Move month_name() in here from App::WRT, add tests.

  - App::WRT::EntryStore
    - Hash file types for entries (directory or flatfile)
    - Use keys of file type hash for complete list of entries.
    - has_prop($entry, $property)
    - is_dir($entry), is_file($entry), is_extant($entry)
    - parent_of($entry)
    - has_index($entry)
    - Make EntryStore cache whether a file is a flatfile or a directory, as
      well as its existence, in a single hash.
    - Include index flatfiles in @source_files for use by has_index()
    - Various tests.

  - App::WRT::FileIO
    - Still duplicates a bunch of shit from Util, so that needs sorted.

  - App::WRT::Renderer
    - Convert to a proper class.
    - Add experimental FileIO class to use in Renderer (imperfect,
      tricky, still thinking about this).  The idea is to separate out the
      concerns of reading and writing the filesystem.

  - App::WRT
    - Refactor display() and improve tests
      - Use "@entries" instead of "@options" for clarity
      - Handle entry names that might evaluate as false
      - Test running display() without any params
    - Rename expand_option() -> expand_alias(), refactor
    - Use EntryStore::has_prop() to detect wrt-noexpand.prop
    - year(), month(), entry() partially rewritten to use EntryStore
    - year() should handle months which are a flatfile
    - Refactor icon_markup() to use is_file() / is_dir() / is_extant(),
      add tests.
    - Add subtitle to feeds

  - bin/wrt-ls is now a "modulino" with tests

  - bin/display errors on non-existent entries

  - Build.PL
    - Remove bogus XML::Feed dependency

wrt-ls is a provisional interface to the tools in WRT::EntryStore.  It
should do more, but work is needed that I'm not going to put in before
releasing v5.0.0.

This seems like a good point to call things stable for this major release.
I have a bunch of ideas that I'd like to implement for a v6, which might
well break the API and underlying data structures.

This is all pretty academic since no one else actually uses this thing.