When debugging the fstring table, I found "UTF-8" to be the most common
interned strings in many benchmarks.

We have a fixed, limited number of these strings, so we might as well permanently
cache their fstrings.

Encodings are RTypedData, not the deprecated RData. Although the
structures are compatible we should use the correct API.

to adopt strict shareable rule.

* (basically) shareable objects only refer shareable objects
* (exception) shareable objects can refere unshareable objects
  but should not leak reference to unshareable objects to Ruby world

Previously we were looping over the enc_table, but when I added an
assertion the only thing that loop was doing is the equivalent of
ENC_TO_ENCINDEX(base). However we don't even need the index of base.

Instead we should be able to just use the badirectly.

Previously we would sometimes call enc_register_at several times in
order to update the encoding values after the base encoding may have
been loaded.

This updates enc_register_at to only be used via enc_register, when an
actual new encoding at a new index is being registered.

Other callers (which in all cases found the index by the name matching)
now call enc_load_from_base which is only responsibly for loading the
encoding from the base values.

Previously when we copied base_encoding on top of the encoding, other
threads could briefly see the name and ruby_encoding_index of the base
encoding.

Not all ractor-related encoding issues were fixed by 1afc07e815051e2f73493f055f2130cb642ba12a.
I found more by running my test-all branch with 3 ractors for each test.