Speed up NDT bucket lookup with a dense hash map

[bloom256]

This PR improves performance of NDT bucket storage and lookup by replacing std::unordered_map with ankerl::unordered_dense::map.

Rationale:

• ankerl::unordered_dense::map is more cache-friendly, faster for lookups, and more memory-efficient than std::unordered_map.
• The container provides different iterator/reference lifetime guarantees, but no code paths were found that rely on iterator or reference stability.

Alternatives considered:

• absl::flat_hash_map was tested but showed lower performance in this use case.
• robin_hood::unordered_flat_map was also evaluated, but its API is incompatible and would require a larger refactor.

Licensing:

• ankerl::unordered_dense is MIT-licensed.

Performance:

• ~10% speedup observed in compute_step_2 compared to the previous version.

Notes:

• This is a data-structure substitution only; algorithmic behavior and the mathematical model are unchanged.

[JanuszBedkowski]

please solve build failed Build Python Bindings on Linux / build (3.12) (pull_request)

[bloom256]

please solve build failed Build Python Bindings on Linux / build (3.12) (pull_request)

Done