proper lvl2dim computation for add/sub

[aartbik]

even though this is not used anywhere yet, the DIA formats require
this correction for the proper inverse (restricted to proper
compressed-range pairs for now)

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Copilot]

Pull Request Overview

This PR corrects the inverse mapping from levels to dimensions for LvlOp::Add and LvlOp::Sub, ensuring proper handling of compressed-range pairs even if not yet used by existing code.

• Adds compile-time checks for the next level specifier to be an identity range
• Implements conditional assignments to recover original dimensions in both Add and Sub cases

Comments suppressed due to low confidence (2)

include/matx/core/sparse_tensor_format.h:259

• The new inversion logic for Add/Sub introduces two distinct branches. Add unit tests to cover both cases where ftype::Expr::cj == ntype::Expr::di and ftype::Expr::di == ntype::Expr::di.

        using ntype = cuda::std::tuple_element_t<L + 1, LvlSpecs>;

include/matx/core/sparse_tensor_format.h:259

• [nitpick] Consider adding a brief comment explaining the mathematical rationale behind these inverse computations for LvlOp::Add and LvlOp::Sub to aid future maintainers.

        using ntype = cuda::std::tuple_element_t<L + 1, LvlSpecs>;

[cliffburdick]

/build