fix(provider/moonshotai): preserve complete raw usage objects (#19623)

Closes #19619

Stacked on #19613; the base will advance with the Moonshot parity stack.

## Summary

- parse Moonshot usage and known token-detail objects with loose schemas
- preserve future provider-specific top-level and nested fields in
`usage.raw`
- keep normalized input, cache, output, and reasoning counts unchanged
- cover normal and streaming responses with regression tests

## Verification

- `pnpm --filter @ai-sdk/moonshotai test:node` (178 passed)
- `pnpm --filter @ai-sdk/moonshotai test:edge` (178 passed)
- `pnpm --filter @ai-sdk/moonshotai type-check`
- `pnpm type-check:full`
- `pnpm check`

## Backports

- v6: applicable after this PR merges
- v5: not applicable because `LanguageModelV2Usage` has no `raw` field