feat: inject anthropic prompt cache control

[shudonglin]

⚠️ 提交说明 / PR Notice

Important

• This description was written and checked for this PR. The code changes were AI-assisted and reviewed locally before submission.

📝 变更描述 / Description

Adds gateway-side Anthropic prompt cache control injection for Claude relay requests.

The relay now adds top-level cache_control only when the client request does not already define Anthropic cache control. The default TTL comes from ANTHROPIC_PROMPT_CACHE_TTL, and callers can override per request with x-anthropic-prompt-cache-ttl. auto selects 1h for evaluation, benchmark, batch, pipeline, and long-running workloads, otherwise 5m.

Gateway-only cache policy headers are stripped before upstream conversion so they do not leak to providers.

🚀 变更类型 / Type of change

• 🐛 Bug 修复 (Bug fix) - 请关联对应 Issue，避免将设计取舍、理解偏差或预期不一致直接归类为 bug
• ✨ 新功能 (New feature) - 重大特性建议先通过 Issue 沟通
• ⚡ 性能优化 / 重构 (Refactor)
• 📝 文档更新 (Documentation)

🔗 关联任务 / Related Issue

• Closes # (none)

✅ 提交前检查项 / Checklist

• 人工确认: 我已亲自整理并撰写此描述，没有直接粘贴未经处理的 AI 输出。
• 非重复提交: 我已搜索现有的 Issues 与 PRs，确认不是重复提交。
• Bug fix 说明: 若此 PR 标记为 Bug fix，我已提交或关联对应 Issue，且不会将设计取舍、预期不一致或理解偏差直接归类为 bug。
• 变更理解: 我已理解这些更改的工作原理及可能影响。
• 范围聚焦: 本 PR 未包含任何与当前任务无关的代码改动。
• 本地验证: 已在本地运行并通过测试或手动验证，维护者可以据此复核结果。
• 安全合规: 代码中无敏感凭据，且符合项目代码规范。

📸 运行证明 / Proof of Work

Local verification passed:

go test ./relay ./relay/channel ./relay/channel/claude -run 'TestApplyAnthropicPromptCacheControl|TestProcessHeaderOverride_PassthroughSkipsAnthropicPromptCachePolicyHeaders|TestConvertOpenAIRequest.*PromptCacheControl'

[shudonglin]

E2E cache validation

Ran on 2026-06-28 from branch feat/anthropic-cache-control-injection:

go test -count=1 ./relay ./relay/channel ./relay/channel/claude -run 'TestApplyAnthropicPromptCacheControl|TestProcessHeaderOverride_PassthroughSkipsAnthropicPromptCachePolicyHeaders|TestConvertOpenAIRequest.*PromptCacheControl|TestClaudeAdaptorE2EInjectsPromptCacheControlAndForwardsUsage'

Result: passed (ok github.com/QuantumNous/new-api/relay, ok github.com/QuantumNous/new-api/relay/channel, ok github.com/QuantumNous/new-api/relay/channel/claude).

The E2E test uses a mock Anthropic upstream, so it validates gateway request/response behavior without spending real Anthropic tokens. It verifies top-level cache_control: {"type":"ephemeral","ttl":"1h"} is sent for long/eval workloads, gateway-only cache policy headers are stripped before upstream, client-supplied cache control is preserved, and Anthropic cache usage fields flow back through response handling.