To make debugging easy, and to improve the performance little bit.

The responsibility moved to caller to avoid confusion. Currently the
function is called from only 2 places, so it is relatively easy to
ensure not to update `gray_list` in the caller.  But the assumption
may change in the future.

`mrb_ary_modify` calls `mrb_write_barrier`, so can cause the same
problem of the past `push`. It is provided for use-level API.

When the array is very big, the simpler `mrb_write_barrier` causes
calling `gc_mark_children` for big arrays repeatedly. That would hinder
performance very badly.

Extract memsize to gem

Those methods are originally CRuby specific.
Co-authored-by: Yukihiro "Matz" Matsumoto <matz@ruby.or.jp>

Those tests succeeds only on some configuration.

The old code compiles only on `MRB_NO_BOXING`.

`mrb_value` may or may not be struct according to configuration.

`C` retrieves a `mrb_value` that refers a class/module.
`c` retrieves a `struct RClass*` pointer to a class/module.

`ObjectSpace.memsize_of_all` takes a class and count memory size of all
instances of the class and its subclasses (if any).

Add ObjectSpace.memsize_of_all

This bug has been there since mruby 1.4.0 (2018-04).

Also avoid `mrb_funcall` to minimize VM recursion.

Memory size of a Fiber is calculated by stack size only in CRuby.

This is enhancement from CRuby's `memsize_of`. We need to change the
CRuby first for the enhancement.

Fixed shift width for `MRB_ENV_SET_BIDX()`

Add ObjectSpace.memsize_of

ref c07f24cd and close #5035