Align fold_haswell, upper_haswell, upper_ice to fold_ice canonical ordering

Use fold_ice's branch ordering as the canonical pipeline (most mature
path from upstream). Apply the same priority sequence to all three
derivative functions for maintainability and consistency.

Canonical pipeline:
    ASCII → Latin-1 (C3) → Cyrillic basic (D0/D1) → Greek (CE/CF)
    → Caseless 2-byte D7-DF (Ice only) → Cyrillic Supplement (D1-ext/D2/D3/D4)
    → Universal 2-byte (Ice only) → 3-byte caseless
    → Nuskhuri/Asomtavruli (E2 B4 ↔ E1 82/83)
    → Mkhedruli/Mtavruli (E1 83 ↔ E1 B2)
    → Latin Extended Additional (E1 B8-BB) → Fullwidth (EF BC/BD)
    → Serial drain

Per-function moves:

  fold_haswell:
    - Swap Cyrillic Supplement ↔ Greek (Greek first).
    - Move 3-byte caseless from tail to after Cyrillic Supplement.

  upper_haswell:
    - Same swap and same 3-byte caseless move.
    - Swap Mkhedruli ↔ Nuskhuri so Nuskhuri runs first.

  upper_ice:
    - Move 3-byte caseless from position 2 (right after ASCII) to
      position 9 (after all 2-byte branches).
    - Move Greek from after Cyrillic Supplement to before it.
    - Swap Mkhedruli ↔ Nuskhuri so Nuskhuri runs first.

Verified bit-perfect haswell-vs-serial and ice-vs-serial across all
1.1M codepoints — 0 diffs.

Bench (Intel Xeon Platinum 8375C, 32 MiB, best of 7):
    ASCII    upper: serial 1.21 → haswell 9.79 (×8.1), ice 9.08 (×7.5)
    Cyrillic upper: serial 0.29 → haswell 3.78 (×13.1), ice 4.50 (×15.5)
    Greek    upper: serial 0.43 → haswell 2.82 (×6.6), ice 3.84 (×9.0)
    Mixed    upper: serial 0.78 → haswell 5.38 (×6.9), ice 6.24 (×8.0)
    CJK      upper: serial 0.29 → haswell 3.27 (×11.2), ice 2.45 (×8.4)

Note: heavily-interleaved scripts (a pattern alternating D2/D3/D4
Cyrillic Supplement leads byte-by-byte) make every chunk advance only
2-4 bytes through the per-lead fast paths, and the per-iteration
pre-checks dominate. This is a pipeline-architecture property, not a
regression — pure D2-only Cyrillic Supplement still runs at 3.5 GB/s
in Ice.

Author: eshishki