fix(billing): price Anthropic 1-hour cache writes separately

[devin-ai-integration]

Summary

Rewrite of #1201 by @atirna (credited as co-author) to match repo conventions; the fix itself is unchanged.

Anthropic reports a TTL breakdown under usage.cache_creation.{ephemeral_5m,ephemeral_1h}_input_tokens; the aggregate cache_creation_input_tokens was priced entirely at the 5-minute 1.25x rate, under-billing 1-hour writes (2x).

• catalog.Pricing gains CacheWriteMultiplier1h (+ DefaultCacheWriteMultiplier1h = 2.0, EffectiveCacheWriteMultiplier1h()).
• catalog.EffectiveInputCost(input, cacheCreation, cacheCreation1h, cacheRead, p, provider): cacheCreation1h is clamped to cacheCreation; the remainder prices at the 5-minute rate, so aggregate-only payloads (Bedrock/Vertex, cacheCreation1h == 0) are unchanged.
• The 1h count is carried from otel.UsageExtractor (SSE message_start + non-stream JSON), translate.UsageSink.RecordCacheCreation1hTokens, handover.Usage.CacheCreation1h and streamCostWriter into billing.DebitInferenceParams.CacheCreation1h, telemetry cost columns, cost headers and policy-outcome cost.

Deltas from #1201: test comments in cost_test.go rewritten as invariant descriptions (no issue refs / pre-fix narrative, per AGENTS.md comment rule); routerCostCalculator / routerResponseCostFromPricing parameter order aligned with EffectiveInputCost (cacheCreation, cacheCreation1h, cacheRead).

Fixes #867

make precommit clean.

Link to Devin session: https://app.devin.ai/sessions/e1c3317bd01c4d7b8a4639c5ef2bae64
Open in Devin Desktop: https://app.devin.ai/desktop/session/e1c3317bd01c4d7b8a4639c5ef2bae64?variant=devin
Requested by: @steventohme