feat(openai): support GPT-5.6 structured output

[bradleybernard]

Description

Adds native structured-output recognition for the official GPT-5.6 API model family:

• gpt-5.6 (Sol alias)
• gpt-5.6-sol
• gpt-5.6-terra
• gpt-5.6-luna

Without these entries, StructuredModeResolver silently falls back to JSON mode, so Prism sends json_object instead of json_schema and ignores schema.strict.

OpenAI documents Structured Outputs support for each GPT-5.6 model:

• https://developers.openai.com/api/docs/models/gpt-5.6-sol
• https://developers.openai.com/api/docs/models/gpt-5.6-terra
• https://developers.openai.com/api/docs/models/gpt-5.6-luna

The regression test covers all four identifiers and asserts Prism sends json_schema with strict: true.

This is intentionally narrower than #991, which proposes broader prefix matching for versioned model names.

Breaking Changes

None.

Test plan

• vendor/bin/pest tests/Providers/OpenAI/StructuredTest.php
• vendor/bin/pest (1467 passed, 8 skipped)
• vendor/bin/phpstan analyse --memory-limit=512M
• Pint and Rector on changed files

[wishborn]

For context from a downstream fork — Particle-Academy/prism already covers this, via prefix matching in StructuredModeResolver rather than an exact-match list. It matches on the gpt-5 family prefix, so gpt-5.6 and the -sol / -terra / -luna variants already resolve to structured mode (sending json_schema with strict) without per-model entries — that's effectively the broader #991 approach this PR intentionally avoids.

Sharing in case it helps the narrow-vs-broad trade-off discussion: the prefix approach is zero-maintenance for new gpt-5.x snapshots, at the cost of assuming every future gpt-5.* supports structured output (which holds for the GPT-5.6 family per OpenAI's docs). The explicit list here is the safer, more deliberate side of that trade-off.