OpenAI prompt caching: parse and report cached token usage

[bug-ops]

Summary

Implement prompt caching metrics reporting for the OpenAI provider, mirroring the existing Claude implementation. OpenAI caches prompts automatically for requests >= 1024 tokens and reports cached token counts in the usage response. We need to parse and surface this data.

Background

The Claude provider already tracks cache usage via:

• last_cache: Mutex<Option<(u64, u64)>> field
• ApiUsage struct with cache_creation_input_tokens / cache_read_input_tokens
• last_cache_usage() trait method on LlmProvider
• record_cache_usage() in Agent aggregates into TUI metrics

OpenAI returns cached token counts in usage.prompt_tokens_details.cached_tokens. No API changes are needed to enable caching -- it is automatic. We only need to parse the response.

Architecture plan

See .local/plan/openai-prompt-caching.md

Tasks

• M0: Project bootstrap — workspace and crate skeleton #1: Add usage deserialization structs and last_cache field to OpenAiProvider
• M1: Ollama chat loop — interactive CLI with LLM #2: Parse cache usage from chat and tool responses
• M2: Skills system — SKILL.md loading and shell execution #3: Implement last_cache_usage() for OpenAI provider
• M3: Persistent memory and Claude provider #4: Add unit tests for usage parsing

Files to modify

• crates/zeph-llm/src/openai.rs

Acceptance criteria

• OpenAiProvider::last_cache_usage() returns (0, cached_tokens) after API calls
• Cache usage is logged via tracing::debug! (matching Claude pattern)
• TUI metrics display OpenAI cached tokens via existing record_cache_usage() path
• All existing tests pass, new unit tests for usage deserialization added
• cargo clippy --workspace -- -D warnings passes