feat: expose raw HTTP responses

[ktamulonis]

Summary

Adds an explicit raw-response modifier for generated HTTP operations:

raw = openai.models.with_raw_response.retrieve("gpt-5.2")

puts(raw.status)
puts(raw.headers["openai-processing-ms"])
puts(raw.request_id)
puts(raw.read)

model = raw.parse

OpenAI::RawResponse[T] exposes the HTTP status, normalized immutable headers,
request ID, buffered body, and a lazy cached parsed value without changing the
operation's arguments or normal error behavior.

API contract

• Ordinary top-level models and pages retain the _request_id convenience
  introduced by feat: expose request IDs #352.
• raw.request_id and raw.parse._request_id are the same value when the
  parsed result is a model or page.
• Raw responses provide metadata for model, page, primitive, binary, and nil
  results.
• Full response headers are not attached to models, pages, nested values, or
  helper-assembled streaming snapshots.
• Higher-level streaming helpers continue to expose status and headers
  directly on their stream objects.
• Generated resource navigation and HTTP method signatures are preserved in
  Ruby, RBI, and RBS.

This follows the explicit raw-response pattern used by the other OpenAI SDKs
while keeping transport metadata separate from API-schema models.

Generator ownership

The generated resource wrappers and type signatures are owned by the Castiron
source change in openai/openai#1272460. That companion PR prevents future SDK
generation from overwriting this API.

Validation

• Ruby 4.0.6: 567 tests, 2,903 assertions
• Ruby 3.3.12: 567 tests, 2,903 assertions
• Raw-response contract: 9 tests, 1,001 assertions
• RuboCop: 2,608 files, no offenses
• Sorbet: no errors
• RBS: 1,214 files validated
• Gem build and package-content inspection
• Thermo-nuclear code-quality review

[HAYDEN-OAI]

Overall looks good. I left one low-priority serialization-boundary comment inline.