Fix concurrent feature memoization

[jnunemaker]

DSL feature lookup now memoizes through a thread-safe map, so shared Flipper instances return one Feature object per feature even when multiple threads populate the cache at once. This replaces non-atomic Hash mutation with Concurrent::Map#compute_if_absent while preserving existing key normalization and feature construction behavior. Covered by a new concurrent DSL spec.

Tested with bundle exec rspec spec/flipper/dsl_spec.rb and git diff --check.

---