…mentation

Replaces the René Nyffenegger header-only base64.h with simdutf v8.2.0,
the same SIMD-accelerated codec used internally by Node.js, Bun, and
Deno. Vendored as the upstream amalgamation (cpp/simdutf.{h,cpp})
keeping the zero-external-dependency setup.

- Encode (base64FromArrayBuffer) calls simdutf::binary_to_base64
  directly into a pre-sized output buffer, with no intermediate
  std::string copy.
- Decode (base64ToArrayBuffer) mirrors V8's Uint8Array.fromBase64
  via simdutf::base64_to_binary_safe with last_chunk_handling::loose
  and decode_up_to_bad_char=true. The decoded std::string is moved
  into a DecodedBuffer (jsi::MutableBuffer) and returned as
  jsi::ArrayBuffer(rt, shared_ptr<...>) — no memcpy.

Roundtrip throughput now matches Hermes' built-in atob/btoa on the
example app's image.json benchmark.

Behaviour changes (TC39-compliant, see Uint8Array.fromBase64 spec):

- Mid-stream padding (e.g. 'SQ==QU0=') now throws instead of
  silently decoding the prefix.
- Embedded '\n' without removeLinebreaks=true now throws instead
  of being silently ignored.

Test fixtures in example/src/tests/{corrupt,linebreaks}.ts updated
accordingly.

Adapted from craftzdog#50, ported onto the post-codegen QuickBase64Impl.cpp
and combined with the MutableBuffer-based no-memcpy decode path.

Closes craftzdog#50

Co-authored-by: Karol Binkowski <karolbinkowski3@proton.me>