remove dependency on sorted_set

[razumau]

sorted_set stopped working in Ruby after internal Set changes (changes description: https://bugs.ruby-lang.org/issues/21216, https://rubyreferences.github.io/rubychanges/4.0.html#set).

Set subclasses cannot use @hash anymore, which breaks the way sorted_set works (it replaces @hash with a red-black tree from rbtree).

We only need to preserve an ordered list of dependencies to build the same cache key every time (introduced in this commit: e132533). So, we can sort them before building a cache key and use a regular Set.

Alternatively, we could rebuild a new Set every time a dependency is added:

def dependencies
  return Set.new if self == Curly::Presenter
  @dependencies ||= Set.new
  Set.new(@dependencies.union(superclass.dependencies).sort)
end

def depends_on(*dependencies)
  @dependencies ||= Set.new
  Set.new(@dependencies.merge(dependencies).sort)

[bquorning]

Given that SortedSet was introduced in order to fix “issues in the cache layer”, I think it makes sense to sort when we compute the (memoized) cache key instead of every time #dependencies is called.

However, we don’t know if any of our users are using #dependencies to calculate their own cache key – so maybe this is a breaking change?

[libo]

Yeah the breaking change issue was why I suggested to make the existing methods behave as before. Are we afraid of the cost of calling sort on every read? It's a O(n log n) where n won't be bigger than ~20 maybe 30...

[libo]

BTW original PR ...summer of 2014 #78