Add MethodDriftTest to catch method drift between RBS and core

[zonuexe]

Stacked on #3218. Only the last 3 commits are new; the test fails without the singleton methods that #3218 adds. I will rebase this onto master after #3218 is merged.

test/stdlib/method_drift_test.rb works like ConstantDriftTest, but checks methods instead of constants. For a fixed list of platform-invariant core classes (Integer, String, Hash, Binding, ...), it checks both directions:

• Missing from RBS: methods that Ruby defines are declared in RBS. The test compares singleton_methods(false) with DefinitionBuilder#build_singleton, and public_instance_methods(false) + protected_instance_methods(false) with #build_instance. Only methods implemented in C or <internal:...> count, so methods added by libraries such as pp and json/add are ignored. A method declared on an ancestor counts as declared.
• Gone from runtime: methods that RBS declares on the class itself are defined at runtime, either public or private. Methods declared with the bot return type are skipped, because RBS uses it for methods that Ruby undefines (e.g. Complex#<).

To check that the test works, I replaced core/hash.rbs with the version from master. The test then reported ::Hash.ruby2_keywords_hash and ::Hash.ruby2_keywords_hash?.

Exceptions in MethodDriftTest::SKIP

Method	Reason
Numeric#+, #-	Declared on Numeric so that code typed as Numeric can add and subtract (aaf7623). Only the subclasses define them
Struct.members, .keyword_init?	Defined only on the classes that Struct.new creates
Enumerable#to_enum, #enum_for	TODO: Defined in Kernel. These declarations are no more precise than Kernel#enum_for, so they may be removed
Method#box	TODO: Added with the experimental Ruby::Box in Ruby 4.0, which has no signature yet

The first two rows are now documented in a new Q&A item in docs/CONTRIBUTING.md, "Why does RBS declare methods that Ruby does not define on that class?", together with the bot convention.

Other changes

• Adds Binding#irb: (?show_code: boolish) -> void. It is defined in <internal:prelude> and was not declared anywhere. It has no test, because calling it starts an IRB session.

Ruby development builds

On a development build of Ruby (RUBY_PATCHLEVEL < 0, i.e. ruby-head), the test reports drift with notify instead of failing. New methods appear on head before they can be declared in RBS: the first CI run of this PR found 22 on Ruby 4.1.0dev, such as String#bit_get and Range#clamp. On Ruby 4.0 the test fails as usual. Whether to also run it as a failing nightly job on head can be decided separately.

[ksss]

Calling a method that returns bot does not mean a type error.
For example, Kernel#raise is a type safe method.

[ksss]

Since adding type definitions and adding test features are separate proposals, it would be helpful if you could split them into separate PRs.

[ksss]

Ah, so this was the stack from #3218. My apologies. The PR had already been split.