fix: read usage from finish chunk's totalUsage instead of custom usage event (#670)

The previous approach awaited result.totalUsage after stream ends and
yielded a custom 'usage' chunk. This is unnecessary since the AI SDK's
finish event already carries totalUsage. Read it directly in the
stream-chunk-processor and remove the custom usage event plumbing.

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>