Intl.DateTimeFormat: emit text without materializing parts

Problem

Intl.DateTimeFormat.prototype.format still builds a structural DtfPart list, allocates each formatted numeric field, optionally performs a second digit-translation pass over those parts, and then concatenates every part into the returned string. That work is necessary for formatToParts and interval analysis, but not for text-only formatting.

This remains after #572 compiled immutable resolved formatter state.

Scope

• Introduce one authoritative DateTimeFormat emission algorithm with direct text and structural-parts sinks.
• Stream text-only formatting directly into the final byte buffer, including non-Latin digit translation as fields are emitted.
• Keep exact part boundaries for formatToParts and range formatting without duplicating locale/calendar/Temporal ordering logic.
• Preserve coercion order, TimeClip errors, Temporal field restrictions, OOM behavior, GC safety, and read-only no-GIL sharing.
• Measure exact-parent text/parts/range controls and retain exact test262 DateTimeFormat and Temporal results.

Out of scope

• New CLDR interval-pattern coverage.
• Constructor resolved-state work (completed in #572).
• Promise/microtask/Worker/job lifecycle work.

Parent: #475 Depends on: #572

Landed on main in perf(intl): stream DateTimeFormat text emission (#573).

The retained implementation runs one ordered DateTimeFormat emitter through direct-text and structural-parts sinks. Text formatting no longer materializes DtfPart records; numeric fields use bounded stack scratch, numbering-system digits are translated during emission, and structural consumers copy ephemeral bytes before publication. formatToParts and range consumers keep their exact boundaries and share the same ordering logic.

Exact-parent evidence at 100 jobs:

• all text/parts/equal-range/equal-range-parts checksums match the parent and system JSC;
• text-consumer backing allocation bytes: 90,434,989 → 80,579,965 (−10.9%);
• GC-cell allocations: 234,716 → 234,716;
• collection count: 8 → 8;
• timing windows were noisy, so no throughput ratio is claimed.

Verification on the committed source:

• DateTimeFormat test262: 244/244;
• Intl402 Temporal: 2029/2029;
• built-ins Temporal: 4603/4603;
• affected test262 flips: 0;
• focused TSan DateTimeFormat group: 6 passed / 0 failed / 0 leaked, including moving GC and four-lane no-GIL sharing;
• full unit gate: 1657 passed / 1 skipped / 0 failed / 0 leaked across 10 shards;
• zig fmt --check and git diff --check: clean.