Refactor JIT LTO kernel generation

[KyleFromNVIDIA]

Add an algorithm that computes a matrix product, and add a generic
CMake function that uses this algorithm to generate a matrix of kernels
with desired parameters.

[dantegd]

Minor comments, otherwise this looks great!

[divyegala]

I think we can normalize how keys are named in the JSON file. If the key ends with:

1. _name: Actual name of the template
2. _type: Always used to instantiate the actual kernel function template
3. _type_tag: Template parameters for registerAlgorithm. The JSON should have the whole tag as a value tag_type_ui.
4. _type_val: Value for the types used in constructor of registerAlgorithm as _name + _type_val

[KyleFromNVIDIA]

I think we can normalize how keys are named in the JSON file.

It's not entirely clear to me what this would look like, or if it would even be possible with the current algorithm. Please post a suggestion with your specific renaming recommendations for interleaved_scan_matrix.json.

[divyegala]

I think the dev team can normalize a naming convention after this PR is merged!

[jameslamb]

I didn't look closely at the source files in src/neighbors/ivf_flat, trusting other reviewers to be much more knowledgeable about those than me.

Left a few comments about the general setup though. Most are minor but one is big... are we SURE it's to require a Python interpreter to build libcuvs?

Leaving a non-blocking "Comment" review, I'll come back and approve once those questions are answered.

[jameslamb]

Approving, but please do consider the other comments I left about the strictness and placement of this Python code.

[KyleFromNVIDIA]

are we SURE it's to require a Python interpreter to build libcuvs?

I already discussed this with @divyegala back in December. One of cuvs's dependencies already requires Python to build, so this isn't introducing a new dependency.

Other projects that want this matrix product algorithm may not have the same luxury. Once CMake 4.3 with its string(JSON) improvements comes out, and once we require CMake 4.3, it will be possible to port the algorithm to CMake script and ditch the Python script.

[KyleFromNVIDIA]

/merge