fix(langchain): Record reasoning tokens as metric

[Dylan Pulver (dylanpulver)]

LangChain's UsageMetadata carries output_token_details.reasoning, and nothing in js/src or
integrations read it. completion_tokens covers the spend, so what was lost is the breakdown:
no way to see how much of a reasoning model's output went to thinking, on any LangChain or
LangGraph span.

Mapped reasoning only. js/src/wrappers/mastra.ts:164-172 settles the scope, picking up the
non-modality detail fields (cacheRead, cacheWrite, reasoning) and leaving the modality ones
alone. OutputTokenDetails also carries audio, image, video, document and text through
ModalitiesTokenDetails; the input side drops its copies too, so those are a separate change.

Zero is recorded, not dropped, matching prompt_cached_tokens: 0 in the existing snapshots.

js/src/instrumentation/plugins/langsmith-plugin.ts:456 has the same gap. Left alone because
#2386 is moving that tree.

Snapshot churn is 30 added lines across 10 files, every one completion_reasoning_tokens, both
sides of each span-tree pair regenerated, no deletions.

Ran pnpm test in js/ (1648 passed), check:typings, lint, and test:e2e:update then
test:e2e once each, not twice. On that assert pass scenarios/git-metadata timed out at 30s
under parallel load. It passes in isolation here in 1.19s and on main in 1.20s, and the two
scenarios this change touches pass, so I read it as local contention. Worth confirming on CI.

Both new unit tests fail against main on assertions.