Add OpenAI usage deserialization structs and last_cache field

[bug-ops]

Parent: #348

Task

Add the necessary structs and field to OpenAiProvider for tracking prompt cache usage.

Implementation

In crates/zeph-llm/src/openai.rs:

1. Add deserialization structs:

#[derive(Deserialize, Debug)]
struct OpenAiUsage {
    #[serde(default)]
    prompt_tokens: u64,
    #[serde(default)]
    completion_tokens: u64,
    #[serde(default)]
    prompt_tokens_details: Option<PromptTokensDetails>,
}

#[derive(Deserialize, Debug)]
struct PromptTokensDetails {
    #[serde(default)]
    cached_tokens: u64,
}

2. Add last_cache: std::sync::Mutex<Option<(u64, u64)>> field to OpenAiProvider

3. Update new(), Clone, and Debug impls for the new field (mirror Claude pattern)

4. Add store_cache_usage() and log_cache_usage() helper methods

Acceptance criteria

• OpenAiUsage deserializes from OpenAI response JSON including prompt_tokens_details.cached_tokens
• last_cache field initialized to None in constructor
• Clone creates independent mutex with None
• Debug output includes cache info