fix(opencode): derive per-model stats from step-finish parts

[Hexecu]

Issue for this PR

Closes #31032

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update

What does this PR do?

opencode stats --models builds its per-model cost and token breakdown by summing message.info.cost and message.info.tokens for each assistant message. Forked sessions copy those message fields verbatim, so the pre-fork history of every fork is counted again in the per-model breakdown.

This changes the per-model aggregation to derive cost and tokens from each assistant message's step-finish parts instead. For a normal session the sum of a message's step-finish part costs/tokens equals message.info by construction (the same usage value is written to both), so the reported numbers are unchanged. Forked sessions zero out their cloned step-finish parts, so deriving from parts stops the breakdown from double-counting pre-fork usage.

This is the per-model breakdown half of #31032. It builds on #31136, which excludes pre-fork usage from the session-level cost/tokens totals by zeroing the cloned step-finish parts on fork. This PR is stacked on #31136, so its diff currently also contains that commit; it should be merged after #31136.

How did you verify your code works?

• Extended the fork cost test in packages/opencode/test/session/session.test.ts to assert that a forked session's cloned step-finish parts carry zero cost and tokens (the values the new aggregation reads), while the parent session keeps its original part cost. bun test test/session/session.test.ts passes (8 tests).
• bun typecheck passes for the opencode package and across the monorepo (22/22).

Checklist

• I have performed a self-review of my code
• I have made corresponding changes to the documentation if needed