[Tracker] Dataset API C++, C, and language wrapper follow-up

Overview

Tracking follow-up items for Dataset API

Sub-tasks:

• [C API] C-based dataset_t needs to store a pointer to a proper C++ Dataset instance.

• Need accessors for matrix shape information

• Need to test these getters in both c and python tests

• [C++ API] Fully support padded dataset as input to CAGRA build

• Verify it works for all of the build algo methods

• Verify this in build kernel code

• [Language Wrappers] Remove dataset views from language wrappers

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3688007585

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3678355568

• @tarang-jain mentioned in slack:
  We should remove the explicit "view" type above and add a member to the dataset_t to designate whether it's owning (this isn't unlike Python ndarray)
  We should make sure the c++ impl of the C apis is properly creating a view when needed (if the dastaset_t passed in is owning)

• Follow up with @aamijar, @tarang-jain, @divyegala

• [Tests] Add test coverage

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3684980300

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3684981518

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3685203633

• [C++ API] Reuse mdarray/mdspan to simplify Dataset API and make it easier to convert between mdarray and datasets

• Address @achirkin ’s C++ changes here: rhttps://github.com/NVIDIA/cuvs/pull/1846#discussion_r3553946083

• [C++ API] More permanent type placeholder to satisfy DatasetViewT template parameter:
  https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3531332959

• [C API][C++ API] Rework file tree organization splitting all dataset classes, structs, and functions into a separate file within include/core/dataset.hpp

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3536786351

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3685314496

• [C++ API] Rename host/device_standard/padded_index type

• Index contains graph + dataset

• Right now naming of index is associated with the type of dataset bound, but should instead be associated with where the graph is located too

• Do we need to introduce a type for the graph? (ex. Host_graph vs device_graph)

• [C API] Migrate merge() Dataset API to have users pass in concatenated dataset + integer offset.

• Migrate merge() the same way we did with extend(). Users should be responsible with concatenating the 2 datasets into 1 merged dataset and provide that single merged dataset + integer offset. Merge() should then merge graph only.

• Think about how to pre-calculate merged storage size after bitset filter:

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3639389310

• [C++ API] Check for hidden H2D copies in ACE path

• Was previously added to preserve / mirror upstream ACE behavior

• Need to re-evaluate because this goes against our contract of never hiding memory allocation from user

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3685764576

• https://github.com/NVIDIA/cuvs/pull/1846#discussion_r3685935639

• [C++ API] Fix update_dataset at C++ API layer

• At C API layer we already centralized an update to a single UpdateDataset function that internally dispatches to AttachDataset or UpdateDeviceDatasetSameLayout based on whether index is standard vs padded

• However, at C++ API layer this can’t be easily done because attach_dataset doesn’t mutate the input index and instead returns a new index whereas update_device_dataset_same_layout mutates the original padded input dataset. C++ has guarantees about immutability for thread safety so we need to revisit how the update_dataset() function can be unified at the C++ API layer.

• [C API][Language Wrappers]

• Cagra-Q path in C API layer and downstream language wrappers was accidentally disabled when compression params was removed.

• Need to expose make_vpq_dataset() factory + rebind compressed dataset on uncompressed index functions at C API and downstream language wrappers. C++ API already has this.