fix(billing): restore GPT-5.6 long-context pricing

[xunyoyo]

Summary

• restore OpenAI GPT-5.6 pricing for requests whose input context exceeds 272K: input-side (including cache read/write) 2x and output 1.5x
• preserve explicit *_above_272k_tokens catalog and channel interval prices instead of clearing or clipping them
• add regression coverage for 272K boundary and 328,375-token cache-read request ($0.338689)

Scope

Forward-only billing behavior. No historical usage records are adjusted.

Verification

• go test ./internal/service -count=1
• go vet ./internal/service
• go test ./... -count=1
• go test -tags integration ./internal/repository/... -count=1

[smallbearrr]

审查发现一个需要修复的问题：

• backend/internal/service/billing_service.go 的 LiteLLM -> ModelPricing 转换路径只传入 base price 和 long-context multiplier，没有把 InputCostPerTokenAbove272KTokens / OutputCostPerTokenAbove272KTokens / cache read/write 的 *_above_272k_tokens 显式高档价格带入实际计费。与此同时，本 PR 的文档和展示层都在保留/展示这些显式 272K+ 价格。

影响：如果远程目录或渠道配置里的显式高档价格不等于 base * 2 / base * 1.5，客户实际扣费会和价格展示/文档描述不一致，可能过扣或少扣。

建议：实际计费路径也应优先使用显式 above-272K 价格，或在转换到 ModelPricing 时把这些高档价格表达出来；同时补一个显式 above-272K 价格故意不等于 multiplier 结果的回归测试，覆盖 input/output/cache read/cache write。

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!