Fix bug for nested require_dependencies

[genkami]

There are problems with nested require_dependencies.

Minimal example that reproduces this bug is: padrino-nested-require.tar.gz

This is how I got this bug:

$ bundle install
$ bundle exec ruby toplevel.rb
  ERROR -  NameError - uninitialized constant ALib:
...
Traceback (most recent call last):
toplevel.rb:10:in `<main>': uninitialized constant ModelWithNoDepOnLib (NameError)
Did you mean?  ModelThatDependsOnLib

Here is how this bug occurs:

• A file toplevel.rb calls require_dependencies to load lib/a_lib.rb, lib/dependent_lib.rb, an_app/app.rb
• The first loading of lib/a_lib.rb fails because it depends on lib/dependent_lib.rb
• lib/dependent_lib.rb are loaded with no error because it does not depend on any other files.
• Then require_dependencies tries to load an_app/app.rb, which may call nested require_dependencies.
• Nested require_dependencies loads an_app/models/model_with_no_dep_on_lib.rb with no error.
• Then nested require_dependencies tries to load an_app/models/model_that_depends_on_lib.rb, but it fails because model_that_depends_on_lib.rb depends on lib/a_lib.rb.
• After all, nested require_dependencies gives up to load files, raising NameError - uninitialized constant ALib
• Outer require_dependencies catches this exception and it assumes that entire process of loading an_app/app.rb fails, although model_with_no_dep_on_lib.rb was loaded and commited successfully.
• Outer require_dependencies rollbacks every constants that is created before preparing an_app/app.rb
• As a result, all constants defined in an_app/models/model_with_no_dep_on_lib.rb are removed, even though they are commited and not to be loaded in next nested require_dependencies.

Modify Gemfile in the example above as follows:

gem "padrino", github: "genkami/padrino-framework", ref: "b5186fe3"

Then execute bundle exec ruby toplevel.rb again. No error may happen.

[namusyaka]

Code change looks good, but I'm not familiar with the reloader component.
I'd like @ujifgc to review this for checking the validity.

[ujifgc]

Looks fine to me.

[ujifgc]

Hey @genkami, thanks for your PR and thorough investigation.

In 338111d I changed the original code to fix your test exactly.

Can you think of another test that would break current implementation?

Related to #2197 #2196

[hirocaster]

added tests(case) #2199

[genkami]

Looks good to me!