numr 0.13.0

Triangular solves take a batch of matrices in one call and behave the same on every backend.

Added

• Batched triangular solves — solve_triangular_lower and solve_triangular_upper take a: [..batch, n, n] with b: [..batch, n, m]. The batch dims of b must equal those of a. Every matrix solves in one call: one kernel launch on CUDA, one dispatch on WebGPU. The result has the shape of b. A mismatched b returns an error that names both shapes. A 2-D a keeps its 1-D and 2-D b forms.
• algorithm::linalg::validate_triangular_solve — the shape rules every backend applies to a triangular solve, with TriangularSolveShape as the resolved batch, order and column count.

Fixed

• CPU triangular solves — a b whose row count differs from the order of a returns ShapeMismatch, and a b of 3 or more dims beside a 2-D a returns an error, as on CUDA and WebGPU. Before, CPU panicked on too few rows, ignored extra rows, and read a higher-rank b as 2-D.
• Triangular solves on strided input — a transposed or offset a or b solves correctly on every backend. Before, CPU panicked, CUDA read the underlying memory in storage order, and WebGPU did not honor a strided or offset a.

Changed

• CUDA triangular solves — a solve with several right-hand-side columns runs in one launch instead of a separate launch sequence per column, and returns without a host synchronization.

Full diff: v0.12.0...v0.13.0