String NFC/NFD normalization and lone surrogate handling

[proggeramlug]

Description

Three related Unicode string gaps (all in test_gap_string_methods):

1. NFC/NFD normalization: "é".normalize("NFC").length returns 2 instead of 1; "é".normalize("NFD").length returns 3 instead of 2. Perry stores strings as UTF-8 bytes and .length returns byte length, not codepoint/grapheme count.

2. isWellFormed() / toWellFormed(): Lone surrogate detection returns wrong results. "\uD800".isWellFormed() returns true (should be false).

3. UTF-16 length semantics: JS .length counts UTF-16 code units, Perry counts bytes. Diverges on any string with characters outside BMP.

These are all manifestations of the same root issue: Perry uses UTF-8 byte-oriented string representation while JS uses UTF-16.

[proggeramlug]

Partially resolved in v0.5.20.

.length now returns UTF-16 code units instead of UTF-8 byte count. All position-based string APIs (charAt, slice, substring, indexOf, lastIndexOf, padStart, padEnd) converted to UTF-16 indexing.

Results:

• "café".length → 4 (was 5)
• "日本語".length → 3 (was 9)
• "😀".length → 2 (was 4)
• NFC/NFD .normalize().length now matches Node

test_gap_string_methods diffs: 4 → 2. Remaining 2 are lone surrogate handling (isWellFormed/toWellFormed) — needs WTF-8 string support, tracked separately.

[proggeramlug]

Two of three sub-issues resolved in v0.5.20:

✅ NFC/NFD normalization length — .normalize("NFC").length now returns correct codepoint count
✅ UTF-16 length semantics — "café".length → 4, "日本語".length → 3, "😀".length → 2

Remaining issue (lone surrogate handling) tracked separately in a new ticket.