feat(llm): add configurable prompt cache TTL with 1-hour Claude variant

[bug-ops]

Summary

• Adds CacheTtl enum (Ephemeral | OneHour) to zeph-llm with serde support and exhaustive match arms
• Extends CacheControl with an optional ttl: Option<CacheTtl> field; omitted on the wire when None/Ephemeral — default output is byte-identical to pre-feature
• Introduces build_cache_control(ttl) helper used at all cache construction sites in cache.rs and replaces the JSON-literal tool hint in mod.rs with a typed path
• beta_header() appends extended-cache-ttl-2025-04-25 when TTL is OneHour
• All 7 split_system_into_blocks call sites and both apply_cache_breakpoint call sites propagate TTL
• New ProviderEntry.prompt_cache_ttl: Option<CacheTtl> field wired through provider_factory.rs; TOML: prompt_cache_ttl = "ephemeral" | "1h"
• Migration step in migrate_claude_provider() — copies "1h", suppresses "ephemeral", idempotent on re-run
• 16 new unit tests covering: enum serde, byte-identity, TTL propagation through system blocks and breakpoints, negative deserialization, migration (survive / suppress / idempotent)

LLM Serialization Gate

This PR touches claude/mod.rs, cache.rs, types.rs, and CacheControl serialization. A live API session test with a stable system prompt is required before merge per project rules.

Test plan

• cargo nextest run --workspace --lib --bins — 8207 passed, 20 skipped
• cargo clippy --workspace -- -D warnings — clean
• cargo +nightly fmt --check — clean
• Live API session test with prompt_cache_ttl = "1h" in config (required before merge)
• Verify anthropic-beta: extended-cache-ttl-2025-04-25 header appears in debug dump
• Verify no 400/422 errors on a multi-turn session

Closes #3096