Add tiled_reduce function with Kernel and ConvertTo abstraction (f16 and f32 impls with hardcoded reduce)

[suri-kumkaran]

Add tiled_reduce function with Kernel and ConvertTo abstraction

Description

This issue establishes the core computational pipeline for multi-vector distances (Chamfer/MaxSim). It introduces the tiled_reduce function alongside the Kernel and ConvertTo abstractions to handle chunked processing and type conversion. To establish a working baseline quickly, the final reduction step will be hardcoded directly within the initial f32 and f16 implementations.

Tasks

• Implement the tiled_reduce function to manage the chunking and routing of vector dimensions, ensuring compatibility with the DistanceFunctionMut trait.
• Define the Kernel abstraction to serve as the interface for hardware-specific distance calculations.
• Define the ConvertTo abstraction to manage type conversions seamlessly within the pipeline.
• Implement the f32 and f16 variants using these abstractions, temporarily hardcoding the reduction logic directly inside these implementations to validate the pipeline.
• Ensure the resulting implementation functions correctly within the standalone API.