fix(stream): emit cache-inclusive input token totals

ccUsageToAiSdkUsage set inputTokens.total to the non-cached input only,
parking cache read/write in the detail fields. The AI SDK convention is
the opposite: total includes cached tokens and noCache carries the
fresh-only remainder. Downstream, totalTokens = input + output feeds
OpenCode's tokens.total, which OpenChamber renders as context usage —
so [CMD] models showed ~0.1% instead of ~5%.

Also pass the parsed cache token counts through to the local cost
calculator (via costUsageFromAiSdkUsage) so fresh input, cache reads,
and cache writes bill at their own rates instead of hardcoding zeros.

Closes #36