fix: split large embedding batches that exceed provider aggregate token limits (#19565)

## Background

Large OpenAI and Azure OpenAI embedding batches failed with HTTP 400
instead of returning all embeddings because requests could exceed the
providers' 300,000-token aggregate limit.

## Root Cause

`embedMany` split requests only by the model's input-count limit. The
reproduction confirmed this produced 2,048- and 130-value requests, with
the first exceeding the aggregate token limit.

## Summary

- Adds an experimental symbol capability in `@ai-sdk/provider-utils` for
embedding-model UTF-8 input byte budgets, leaving the versioned
`EmbeddingModelV4` specification unchanged.
- Configures conservative 300,000-byte budgets for OpenAI and Azure
OpenAI.
- Splits embedding inputs in one pass by both input count and byte
budget while preserving per-chunk retries, telemetry, and response
metadata.
- Propagates the experimental capability through embedding middleware
wrappers.
- Documents the batching behavior and includes patch changesets.

## Testing

Added regression coverage for byte-only splitting, exact boundaries,
combined limits, unlimited budgets, oversized single inputs, middleware
capability propagation, and the OpenAI and Azure OpenAI configured
budgets.

## End-to-end Validation

- `pnpm -C examples/ai-functions exec tsx
src/reproduction/issue-10082-embed-many-token-limit.ts` succeeded
against OpenAI, returning all 2,178 embeddings without the
aggregate-token-limit error.

## Related Issues

Fixes #10082

Closes #17073

---------

Co-authored-by: ai-sdk-factory <308175966+ai-sdk-factory@users.noreply.github.com>
Co-authored-by: Gregor Martynus <39992+gr2m@users.noreply.github.com>
Co-authored-by: eposha <54452472+eposha@users.noreply.github.com>