Remove `LEVELS()` accessor

[lionel-]

Part of https://github.com/r-lib/rlang/issues/1706

LEVELS() was added in https://github.com/r-lib/rlang/pull/1187

Used to detect non-utf8 strings in r_obj_encode_utf8() and optimise the default case of no encoding conversion needed:

• Avoid cloning a STRSXP if no translation is needed.
• Avoid allocating a CHARSXP if no translation is needed (in case of mixed encodings in the vector, less important I guess?).

We could just call Encoding(x) <- "UTF8" from R. This will unconditionally clone the input vector unless it doesn't have any references. To preserve our optimisation, either of these would work:

• Add a predicate to the C API to determine if a CHARSXP is encoded in UTF8

• Improve do_setencoding() to only duplicate if needed so that it's a noop in the common case. Bonus points if exported on C side, e.g. as Rf_EnsureUtf8()?

My sense is that these new unconditional allocs would be bad for performance in dplyr. @DavisVaughan you added these optimisations for vctrs, could you confirm please?

[lionel-]

There's now https://github.com/wch/r-source/commit/0c753e41fc06bf5a8825657a5843104a93f28c12 so we should be able to preserve our optimisations.