Split modules and shard EvaluationSession state

[hardbyte]

Summary

• Splits src/lib.rs into private implementation modules with root-level reexports, preserving the existing use gatehouse::{...} API shape.
• Refactors EvaluationSession internals from global source/cache/in-flight maps into per-fact-type FactState<K> with striped cache and in-flight state.
• Keeps FactSource::load_many() outside locks and preserves the cancellation invariant: cancelled loaders cache LoaderCancelled and wake waiters.
• Documents that checker/combinator batch evaluation remains sequential; parallel work is currently policy/fact-source internal unless a future checker-level API makes it explicit.
• Adds contention/cancellation tests and a hot in-RAM benchmark comparing a coarse cached reference against the sharded session path.

Addresses #21.
Addresses #25.

Notes

This PR is stacked on brian/bulk-authz-batch because both issues depend on the v0.3 EvaluationSession / FactSource work in PR #18. Issue #21's proposed Lookup* module contents are not present because PR #18 removed the unconsumed lookup draft; no lookup API is introduced here.

Validation

• cargo fmt --all --check
• cargo build
• cargo clippy --all-targets --all-features -- -D warnings
• cargo doc --all-features --no-deps
• cargo test --doc
• cargo test --all-targets --all-features
• cargo bench --bench permission_checker --no-run
• cargo check --example axum
• cargo check --example pg18_bulk_rebac
• cargo run --example in_ram_rebac
• DATABASE_URL="host=localhost port=15432 user=postgres password=test dbname=awa_test" cargo run --example pg18_bulk_rebac --release

PG18 release example completed against local PostgreSQL 18.3; largest row: 10000,8000,807.565,11.745,6000,x68.8.