feat(bridge): record prompt-cache breakdown on llm_inference events (bridge-v0.3.3)

[New1Direction]

Why

korgex now records a provable prompt-cache breakdown on every llm_inference event (korgex 0.25.0). The local-journal and HTTP transports carry it, but the PyO3 bridge dropped it: korgex's KorgBridgeClient already forwarded cache_read_tokens / cache_creation_tokens / uncached_input_tokens (with a TypeError fallback), but the Rust record_llm_call didn't accept them — so a cache hit was unprovable from a bridge-written journal. This closes that gap.

What

• record_llm_call takes three new optional params and folds them onto the event args only when caching is active (a read or a write happened), mirroring korgex's src/korg_ledger.py::_llm_call_args byte-for-byte.
• A cold turn keeps the legacy two-field args, so older readers and the hash-chain over historical events are undisturbed. Field order is irrelevant — args canonicalize with sorted keys at hash time, the same spec all three independent verifiers use.
• Bump bridge-v0.3.2 → v0.3.3 (Cargo.toml + pyproject + __version__).
• No korgex-side change needed — the breakdown flows through once a 0.3.3 wheel is installed.

Test plan

• cargo check -p korg-bridge clean off a fresh main base
• rebuilt via maturin develop; 19 bridge tests pass, including:
  • new test_record_llm_call_cache_breakdown (warm breakdown lands on args)
  • new test_record_llm_call_cache_creation_only (cache-write-only still records)
  • the existing test_bridge_chain.py suite — so the new field verifies in-chain
• docs: bridge-v0.3.3 CHANGELOG entry + korg-bridge README note

Notes

This branch was cut clean off main and contains only the bridge crate + its docs — no unrelated workspace WIP.