fix: isolate buffer normalization from native layout

[goblinmode2700]

What changed

• Normalize exact bytearray and C-contiguous memoryview values through the existing Python default-hook membrane.
• Preserve pre-copy size checks, hook ordering, BufferError for non-contiguous views, and refusal/hook behavior for binary subclasses.
• Remove the new native buffer-copy paths, leaving src/encode.rs identical to v0.3.0b4 apart from the required non-string mapping-key guidance.

Why

The first publication-disabled 0.3.0b5 rehearsal established a regression in the unrelated control-nested-tag-decode@512 endpoint: +8.6%, simultaneous Bonferroni interval [+7.7%, +9.5%]. The Issue 13 behavior is cold-path normalization and does not require changing native decode code layout, so this follow-up isolates it at the existing Python membrane.

User impact

Exact bytearray and contiguous memoryview inputs encode as padded base64 before a user enc_hook. Invalid mapping keys still refuse conversion, but the root and nested errors now explicitly name msgspec.to_builtins(..., str_keys=True).

Validation

• make qualify: 488 Python tests passed, 11 skipped; 43 Rust tests passed; official TOON corpus 538/538; G2 zero-intermediate-container checks passed.
• Focused CPython 3.13 ABI3 public-API tests and a sparse 4 GiB pre-copy rejection check passed.
• One preregistered same-machine guard against the public 0.3.0b4 wheel: 36 randomized pairs, 22 endpoints, R 4.4.2 paired-log model, simultaneous Bonferroni intervals.
• Former regression endpoint: +0.5%, interval [-0.4%, +1.5%], non_inferior.
• Family: INCONCLUSIVE, zero regression endpoints. Three unrelated endpoints were inconclusive because only their upper simultaneous bounds crossed 3%; no data-dependent rerun was performed.