fix: preserve prompt cache breakpoints on scalar OpenAI Responses tool results (#19927)

## Background

OpenAI Responses silently discarded explicit prompt cache breakpoints
from scalar tool results, preventing those tool-result boundaries from
participating in explicit prompt caching.

## Root Cause

The Responses scalar-result serializers returned or combined strings
without consistently reading output-level provider options or the
tool-result-level fallback. The immutable reproduction and focused path
coverage confirmed breakpoint loss across ordinary function, custom
tool, and expanded-parallel conversions.

## Summary

Scalar text, JSON, error-text, error-JSON, and execution-denied results
now resolve output-level breakpoints with tool-result-level fallback and
emit input_text arrays when marked. Existing unmarked strings,
output-schema encoding, warning behavior, and parallel ordering are
preserved, and a patch changeset documents the correction.

## Testing

Regression coverage verifies all five scalar output types at both
provider-option placements across ordinary function, custom tool, and
expanded-parallel paths, including output-schema serialization,
warnings, and unchanged unmarked output. OpenAI Node and Edge suites
each passed 972 tests, OpenAI-compatible Node passed 261 tests, and full
type, format, and lint checks passed.

## End-to-end Validation

- `replay_original_reproduction` — exited 0 without the immutable bug
signal.
- `pnpm -C examples/ai-functions exec tsx -e "<live scalar breakpoint
request>"` — the live GPT-5.6 Responses API returned `finishReason:
"stop"`.

## Related Issues

Fixes #19921

Closes #19922

---------

Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com>
Co-authored-by: alvarosevilla95 <1376447+alvarosevilla95@users.noreply.github.com>
Co-authored-by: ai-sdk-factory[bot] <305873210+ai-sdk-factory[bot]@users.noreply.github.com>
Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>