[Hash] Update Hash's signatures and tests to be correct

[sampersand]

PHEW! This was a long one, but now Hash is finally done!

This PR updates all of the definitions within hash.rbs to be more correct, as well as modernizes the entire Hash_test.rb file, which was still using the old format.

A complete list of functional changes is below. Bolded ones are important.

• Hash::_Pair[K, V] is added; Hash.[] and Hash#to_h use it.
• Hash::_Equals is added. Used by Hash#{assoc,has_value?,rassoc}.
• Hash.[] has been changed to return instances in the Hash[a: 1] and Hash[[:a, 1]] cases, and now uses _Pair
• Hash.try_convert returns nil in the untyped case.
• Hash#{<,>,<=,>=} all now accept hash[untyped, untyped]
• Hash#[] now accepts Hash::_Key
• Hash#any?'s signature has been fixed: the block form actually is a tuple, and now uses Enumerable::_Pattern
  • It would be nice to eventually change Enumerable::_Pattern to take a generic
• Hash#{assoc,rassoc,has_value?,key} now all take Hash::_Equals
• Hash#compact returns an instance. There's no real good way to represent the return value for the function, but instance is more accurate than Hash[K, V]
• Hash#deconstruct_keys now has an untyped argument, as it's ignored
• Hash#{default,dig,delete,except,fetch,fetch_values,has_key?,slice,values_at} now all accept Hash::_Key, and any related blocks also do as well
• Hash#default_proc now returns a (^(self, _Key) -> V)?. I'm not sure if this self is legal, or needs to be an instance (or even Hash[K, V], though I don't like that because it doesnt account for Hash subclasses)
• Hash#default_proc= is similarly plagued with Hash#default_proc's concern. Also, the three branches were split up.
• Hash#each{,_key,_value}'s blocks now returns void not untyped
• Hash#flatten now explicitly handles a level of 0, and now accepts ints
• Hash#{merge,merge!} now both accept hash. Unfortunately, for merge return type is Hash, but really it should be "instance[A | K, B | V]"
• Hash#freeze has been added
• Hash#replace now accepts hash
• The alias pairs size/length, filter/select, and filter!/select! have been swapped, to more closely match documentation. (Other aliases which matched the documentation haven't been touched)
• Hash#to_h's block form can now return _Pair[A, B]
• Hash#{transform_keys{,!}} variants with hash arguments ({a: 1}.transform_keys(a: :b) #=> {b: 1}) have been added
• Hash#initialize's proc now accepts an instance

[sampersand]

This is the exact same snippet as https://github.com/ruby/rbs/pull/2683/files#diff-b00b4cb204913354ffdc33b64788f911df698b99c4c65bf6cfa6bbea17e2870a... If either that or this PR is merged, i'll rebase the other

[sampersand]

Pushed a new version which retains %a{implicitly-returns-nil} on []. removing that should be another PR.

[soutaro]

I'd keep the base case for situations that static type information of the argument is lost, but actually is a to_hash compatible object.

Suggested change

	| (untyped) -> nil
	| (untyped) -> Hash[untyped, untyped]?

[soutaro]

This is technically correct but controversial.
I’m assuming it makes more sense to get a type error when we pass an unexpected value.

hash = { foo: :bar } #: Hash[Symbol, Symbol]
hash["foo"]          # This is usually a mistake.

[soutaro]

Oops... rake typecheck_test is not working...

I think having a few test cases especially around _Pair makes sense, because we can confirm the updated type definition works with Steep and test if unexpected new type errors will be reported. (But it's not working now.)