Ref: #402
Ref: https://bugs.ruby-lang.org/issues/10222
Ref: ruby/ruby@5754f15
Ref: ruby/ruby@b6d3927

Up to Ruby 2.3, `require` would resolve symlinks, but `require_relative` wouldn't:

```ruby
require 'fileutils'
FileUtils.mkdir_p("realpath")
File.write("realpath/a.rb", "p :a_loaded")
File.symlink("realpath", "symlink") rescue nil

$LOAD_PATH.unshift(File.realpath(__dir__) + "/symlink")
require "a.rb" # load symlink/a.rb in 2.3 and older, load realpath/a.rb on 2.4 and newer
require_relative "realpath/a.rb" # noop on 2.4+
```

This would easily cause double loading issue when `require` and `require_relative`
were mixed, but was fixed in 2.4 (https://bugs.ruby-lang.org/issues/10222).

The problem is that `Bootsnap` kinda negated this fix, because `realpath()`
wouldn't be applied to absolute paths:

```ruby
require 'fileutils'
FileUtils.mkdir_p("realpath")
File.write("realpath/a.rb", "p :a_loaded")
File.symlink("realpath", "symlink") rescue nil

$LOAD_PATH.unshift(File.realpath(__dir__) + "/symlink")
require File.expand_path("symlink/a.rb") # load symlink/a.rb in 3.0 and older, load realpath/a.rb on 3.1 and newer
require_relative "realpath/a.rb" # noop on 3.1+
```

And for performance reasons, Bootsnap tried really hard not to call `realpath`,
as it's a syscall, instead it used `expand_path`, which is entirely in use
space and doesn't reach to the file system. So if you had a `symlink` in
`$LOAD_PATH`, `bootcsnap` would perpetuate this bug, which led to the
addition of #136.

This was ultimately fixed in Ruby 3.1 (https://bugs.ruby-lang.org/issues/17885),
now `realpath` is applied even on absolute paths.

While `realpath` is indeed expensive, I think the performance impact is ok if
we only call it for `$LOAD_PATH` members, rather than for all requirable files.
So if you have X gems, it's going to be more or less X `realpath` calls.

It would stay a problem if a gem actually contained symlinks and used
`require_relative`, but it's quite the stretch, and with 3.1 now
handling it, it's not worth keeping such workaround.

See: #402

Drop Ruby 2.3 support.

Ref: #402
Ref: https://bugs.ruby-lang.org/issues/10222
Ref: ruby/ruby@5754f15
Ref: ruby/ruby@b6d3927

Up to Ruby 2.3, `require` would resolve symlinks, but `require_relative` wouldn't:

```ruby
require 'fileutils'
FileUtils.mkdir_p("realpath")
File.write("realpath/a.rb", "p :a_loaded")
File.symlink("realpath", "symlink") rescue nil

$LOAD_PATH.unshift(File.realpath(__dir__) + "/symlink")
require "a.rb" # load symlink/a.rb in 2.3 and older, load realpath/a.rb on 2.4 and newer
require_relative "realpath/a.rb" # noop on 2.4+
```

This would easily cause double loading issue when `require` and `require_relative`
were mixed, but was fixed in 2.4 (https://bugs.ruby-lang.org/issues/10222).

The problem is that `Bootsnap` kinda negated this fix, because `realpath()`
wouldn't be applied to absolute paths:

```ruby
require 'fileutils'
FileUtils.mkdir_p("realpath")
File.write("realpath/a.rb", "p :a_loaded")
File.symlink("realpath", "symlink") rescue nil

$LOAD_PATH.unshift(File.realpath(__dir__) + "/symlink")
require File.expand_path("symlink/a.rb") # load symlink/a.rb in 3.0 and older, load realpath/a.rb on 3.1 and newer
require_relative "realpath/a.rb" # noop on 3.1+
```

And for performance reasons, Bootsnap tried really hard not to call `realpath`,
as it's a syscall, instead it used `expand_path`, which is entirely in use
space and doesn't reach to the file system. So if you had a `symlink` in
`$LOAD_PATH`, `bootcsnap` would perpetuate this bug, which led to the
addition of #136.

This was ultimately fixed in Ruby 3.1 (https://bugs.ruby-lang.org/issues/17885),
now `realpath` is applied even on absolute paths.

While `realpath` is indeed expensive, I think the performance impact is ok if
we only call it for `$LOAD_PATH` members, rather than for all requirable files.
So if you have X gems, it's going to be more or less X `realpath` calls.

It would stay a problem if a gem actually contained symlinks and used
`require_relative`, but it's quite the stretch, and with 3.1 now
handling it, it's not worth keeping such workaround.

See: #402

Get rid of RealPathCache and the `Kernel.require_relative` decorator

The typo was introduced by accident in #399.

Fix typo

Kernel#require turns its argument into a path by calling `to_path` if it
responds to it, and then into a String (the argument or the return value
of `to_path`) by using implicit conversion (`to_str`).

Co-authored-by: Étienne Barrié <etienne.barrie@gmail.com>

Bind `rb_get_path` to better respect Kernel#require duck-type

Fix: #404

But also it's annoying to have delayed `require` like this one.

Drop dependency on FileUtils

Fix: #411

The load path cache can be mutated, so we need to unfreeze the hash
when it happens.