P2728: loosen wording to admit chunked (SIMD) implementations

[ednolan]

The wording previously pinned the transcoding iterator to decoding one code point per read: buf_ was sized to hold exactly one transcoded code point, and base() and iterator equality were specified directly in terms of the exposition-only members, making the buffer anchor observable and blocking an as-if chunked implementation. Loosen it so an implementation may transcode chunks of input at a time, e.g. with SIMD:

• Make buf_'s capacity an unspecified constant, buffer-capacity, of at least 4 / sizeof(ToType), and widen buf_index_ and to_increment_ accordingly. Chunking is permitted only when the underlying range models forward_range; read-ahead on a single-pass range is destructive and therefore observable.
• Allow read() to transcode an implementation-chosen number n >= 1 of consecutive input subsequences per invocation, provided the result fits in buf_. The choice of n is unobservable and may vary call to call: an implementation typically transcodes a fixed-size window of code units trimmed back to whole input subsequences. Substitution of Maximal Subparts applies per input subsequence exactly as before.
• Allow read-reverse() to chunk symmetrically (n subsequences ending at current_), specify explicitly that it leaves current_ at the beginning of the first subsequence it transcoded, and note that the chunk's starting boundary is locatable by bounded backward scanning.
• Respecify base() and iterator equality positionally -- in terms of the position of the current input subsequence and the offset of the current element within its transcoded code units -- instead of by memberwise comparison, since iterators denoting the same element can hold different buffer anchors once buffers are chunked.
• Make operator-- skip an ill-formed input subsequence's three-unit replacement-character encoding as a unit in _or_error views with char8_t output, mirroring operator++: with chunked buffers those code units can sit mid-buffer, and the skip preserves the canonical first-unit position that positional equality relies on.
• Add a "SIMD support" design discussion section: iterator size and ABI implications of the buffer capacity, base() implementation strategy, validation-plus-scalar-fallback for SMS error handling, interactivity considerations, preliminary UTF-16 to UTF-8 performance numbers from the prototype std::simd kernel (enolan_simd4 branch), and why a view cannot approach bulk transcoding speed. Update the changelog.

[coveralls]

coverage: 99.744%. remained the same — enolan_simdwording2 into main