feat: add llm-types feature gate for assertion types without reqwest

[noahgift]

Summary

The llm feature gate currently requires reqwest, tokio, and futures as dependencies. However, the assertion types (ChatResponse, LlmAssertion, LlmAssertionResult, assert_deterministic) are pure Rust structs + validation logic with zero network dependencies.

Downstream consumers like apr-model-qa-playbook need the typed response validation but already have their own HTTP layer (CommandRunner trait). Importing llm forces 3 heavyweight deps that aren't used.

Proposed Change

Split the llm module into two feature gates:

• llm-types — exports client types (ChatResponse, ChatMessage, Usage, Role, etc.), assertion module (LlmAssertion, assert_deterministic), and report module. Zero new deps.
• llm — depends on llm-types + adds reqwest/tokio/futures for LlmClient and LoadTest.

Motivation

apr-qa-runner serve battery (19 endpoint checks) validates OpenAI-compatible JSON responses. Currently reimplements schema validation, determinism checks, and token count assertions inline. With llm-types, it can use jugar_probar::llm::assertion::LlmAssertion directly.

Files Changed

• crates/probar/Cargo.toml — add llm-types feature
• crates/probar/src/llm/mod.rs — conditional compilation for client HTTP methods
• crates/probar/src/llm/client.rs — split types (always) from HTTP impl (llm only)

[noahgift]

Triage: P3 — WASM testing feature. Part of the multi-threaded WASM UX testing epic (#14). Not blocking core functionality.