fix: send model-specific OpenAI reasoning formats for Bedrock Converse (#19420)

## Background

CRIS-prefixed OpenAI Converse requests mapped reasoning effort to
Nova-style `reasoningConfig`, causing Bedrock to reject the request with
HTTP 400 `unknown_parameter`.

## Root Cause

Two model-family assumptions caused the failure:

- OpenAI detection used `modelId.startsWith('openai.')`, which excluded
`us.` and `global.` CRIS profile IDs.
- The OpenAI branch assumed every OpenAI Bedrock model used flat
`reasoning_effort`, but GPT-5.x requires nested `reasoning.effort` while
gpt-oss requires the flat field.

## Summary

- Recognize direct OpenAI model IDs and IDs with one CRIS profile prefix
without broadly matching arbitrary custom IDs.
- Send nested `reasoning.effort` for GPT-5.x models.
- Preserve flat `reasoning_effort` for gpt-oss models.
- Add a patch changeset describing the model-specific serialization fix.

## Testing

- Added request-body regression coverage for `us.openai.gpt-5.6-luna`
and `global.openai.gpt-5.6-luna`, asserting nested `reasoning.effort`
and the absence of both rejected fields.
- Preserved the existing gpt-oss flat-field regression coverage.
- Added coverage ensuring arbitrary custom model IDs containing
`openai.` are not reclassified.
- `pnpm --filter @ai-sdk/amazon-bedrock test` — 19 Node and 19 Edge test
files passed, 474 tests in each runtime.
- `pnpm --filter @ai-sdk/amazon-bedrock type-check`
- `pnpm check`
- `pnpm type-check:full`

## Service Validation

Live Bedrock validation reported in #19403 confirmed that
`reasoning.effort` succeeds for the `us.` and `global.` GPT-5.6
inference profiles, while flat `reasoning_effort` is rejected. The
gpt-oss family continues to honor the flat field.

## Related Issues

Fixes #19403

Closes #19410

---------

Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com>
Co-authored-by: mmurilo <43329254+mmurilo@users.noreply.github.com>
Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>