feat: add provider-aware prompt caching

[devproje]

What this changes

Provider configuration previously assumed an OpenAI-compatible API and only recorded cache reads exposed through cached_tokens. This meant native Anthropic endpoints could not be used, Claude prompt caching could not be explicitly configured, and cache writes were invisible.

This change:

• adds openai and anthropic provider kinds
• adds auto, off, ephemeral, and ephemeral_1h prompt-cache policies
• supports the native Anthropic Messages API, including streaming, reasoning, images, tool calls, and cache usage
• enables automatic prompt caching for native Anthropic and Claude models routed through OpenRouter
• adds opt-in OpenRouter whole-response caching with a configurable TTL
• records cache reads and writes separately
• adds the cache_write_tokens database migration
• exposes cache-read and cache-write totals through CLI and TUI usage output
• preserves compatibility with existing provider configurations
• documents provider and cache configuration, including response-cache retention caveats

How it was verified

• make test-race passes
• New behaviour has a test, or there is nothing to test

Added coverage for:

• OpenAI-compatible top-level cache control
• automatic OpenRouter Claude prompt caching
• native Anthropic streaming and cache usage normalization
• rejection of response caching on non-OpenRouter endpoints
• cache-write token persistence and aggregation
• migration of existing databases

Checklist

• Follows docs/CONVENTION.md: no comments, no :=, one var block per function in first-use order with err last, callees before callers, and main unconditionally last
• New .go files carry the two-line SPDX header
• Documentation updated if behaviour a user can see has changed
• My contribution is licensed GPL-3.0-or-later, matching the project