Improve internal trie logic

[Longarithm]

I plan to include here all housekeeping issues related to Trie impl. In theory it should be exciting to improve the code and make it more maintainable. Even if we switch to AVL or other tree over state, from Trie we can get some insights on where complexities of the code are hidden.

• refactor: no logic duplication in TrieRecordingStorage #9003
• fix: visit_nodes_interval should exclude path_end #9043
• Explicitly expose trie metrics to TrieStorage callers #9004
• Cleanup trie state parts logic #1603
• Strongly typed trie keys #8998
• Simplify trie iteration - get rid of Crumbs and IterSteps
• Consider eliminating logic duplication in validate_state_part and apply_state_part
• "Refcount leaks" during applying state parts. See also GC + shard rotation leaks #2569.
• Get rid of as_caching_storage, etc: refactor: remove majority of as_caching_storage calls #9006
• Most of state parts tests are based on random data which is not great.