unified interface for all index types

[BlaiseMuhirwa]

We should use a type-erased interface, such as the following (perhaps with reduced virtualization).

/**
 * Type-erased interface for the index class. 
 */
class PyIndexBase : public std::enable_shared_from_this<PyIndexBase> {
protected:
  typedef std::pair<py::array_t<float>, py::array_t<int>> DistancesLabelsPair;

public:
  virtual ~PyIndexBase() = default;

  virtual void add(const py::array &data, int ef_construction,
                   int num_initializations, py::object labels = py::none()) = 0;
  virtual std::shared_ptr<PyIndexBase> allocateNodes(
      const py::array_t<float, py::array::c_style | py::array::forcecast>
          &data) = 0;
  virtual DistancesLabelsPair searchSingle(const py::array &query, int K,
                                           int ef_search,
                                           int num_initializations) = 0;

  virtual DistancesLabelsPair search(const py::array &queries, int K,
                                     int ef_search,
                                     int num_initializations) = 0;

  virtual void save(const std::string &filename) = 0;

  virtual void buildGraphLinks(const std::string &mtx_filename) = 0;

  virtual std::vector<std::vector<uint32_t>> getGraphOutdegreeTable() = 0;

  virtual void reorder(const std::vector<std::string> &strategies) = 0;

  virtual void setNumThreads(uint32_t num_threads) = 0;

  virtual uint32_t getNumThreads() = 0;

  virtual uint64_t getQueryDistanceComputations() const = 0;
};

Then, in the bindings, we can do something like this

struct BindAllDataTypes {
  explicit BindAllDataTypes(
      py::class_<PyIndexBase, std::shared_ptr<PyIndexBase>> &index_class)
      : _index_class(index_class) {}

  template <DataType data_type> void operator()() {
    bindSpecialization<SquaredL2Distance<data_type>, int>(_index_class);
    bindSpecialization<InnerProductDistance<data_type>, int>(_index_class);
  }
  py::class_<PyIndexBase, std::shared_ptr<PyIndexBase>> &_index_class;
};

BindAllDataTypes binder(index_class);
flatnav::util::for_each_data_type<
    BindAllDataTypes, DataType::float32, DataType::int8,
    DataType::uint8>::apply(std::forward<BindAllDataTypes>(binder));

Currently, this doesn't quite work because pybind will bind methods for only the first type (in this case, DataType::float32). When it gets to the next type, DataType::int8, it will notice that the index class already has methods bound, such as add, search, etc. We need to find a way to do something along these lines.