fix(provider/xai): correct usage token calculation for reasoning models

[dancer]

background

follow-up to #11536 — xAI's API reports completion_tokens as text-only, with reasoning_tokens separate in completion_tokens_details (unlike OpenAI which includes reasoning in completion_tokens). per xAI docs: "reasoning token consumption will be counted separately from completion_tokens, but will be counted in the total_tokens"

this results in negative textTokens and underreported outputTokens for all reasoning models (grok-4, grok-3-mini, grok-3-mini-fast, grok-code-fast-1, etc.)

summary

• change outputTokens.total from completion_tokens to completion_tokens + reasoningTokens
• change outputTokens.text from completion_tokens - reasoningTokens to completion_tokens
• rewrite unit tests with real xAI API token values

verification

before (negative tokens)

--- grok-3-mini ---
raw: completion_tokens=2, reasoning_tokens=317, total_tokens=331
sdk: outputTokens=2, textTokens=-315, reasoningTokens=317, totalTokens=14

after (correct)

--- grok-3-mini ---
raw: completion_tokens=1, reasoning_tokens=386, total_tokens=399
sdk: outputTokens=387, textTokens=1, reasoningTokens=386, totalTokens=399

checklist

• tests have been added / updated (for bug fixes / features)
• documentation has been added / updated (for bug fixes / features)
• a patch changeset for relevant packages has been added (run pnpm changeset in root)
• i have reviewed this pull request (self-review)

[dancer]

@Jaaneek can I get your 👀 on this as well please

cc: @shaper will wait on gateway team before merge just to quadriple check

[dancer]

also @williamlmao i know you reported the original issue so tagging you in would love some community feedback once this is merged! pls!

[Jaaneek]

Will take a look in few hours, thanks for working on this

[Jaaneek]

Just confirmed this with the team, will add it to docs on our side as well. Thank you! Good to merge

[vercel-ai-sdk]

⚠️ Backport to release-v5.0 created but has conflicts: #12441