Fix LangChain JS token total metrics

[Stephen Belanger (Qard)]

Summary

• Emit canonical metrics.tokens for LangChain JS LLM spans.
• Prefer explicit LangChain totals when present.
• Synthesize tokens from prompt and completion counts when LangChain omits an explicit total.
• Preserve existing cache token mappings.

Root Cause

LangChain JS callback spans were already logging prompt and completion token metrics, but the callback handler only emitted the legacy total_tokens field for totals. Braintrust cost/token insights expect canonical metrics.tokens, so spans with complete prompt/completion data could still look like they were missing total token usage.

Validation

• mise exec -- pnpm --dir js exec vitest run src/wrappers/langchain/callback-handler.test.ts
• mise exec -- pnpm --dir js exec vitest run src/instrumentation/plugins/langchain-plugin.test.ts
• mise exec -- pnpm exec prettier --check js/src/wrappers/langchain/callback-handler.ts js/src/wrappers/langchain/callback-handler.test.ts .changeset/langchain-js-token-total-metrics.md
• git diff --check

[Luca Forstner (lforst)]

What's the actual diff between tokens and total_tokens?