Make oneMKL sparse matrices AbstractGPUSparseArrays

[michel2323]

Fixes #627.

oneSparseMatrixCSR, oneSparseMatrixCSC and oneSparseMatrixCOO now subtype GPUArrays' AbstractGPUSparseMatrixCSR/CSC/COO, so the generic sparse functionality of GPUArrays.jl applies to them: broadcasting (zero-preserving functions keep the result sparse), sum/mapreduce along dimensions, norm/opnorm, findnz, triu/tril/kron, iszero, scalar indexing under @allowscalar, similar/copy/collect. The GPUArrays sparse testsuite is enabled for the CSR and CSC types (COO is excluded because GPUArrays' generic sparse broadcast only supports vectors, CSR and CSC).

Changes

• Lazy oneMKL handles. GPUArrays' generic code constructs sparse matrices through the plain 4-argument constructor and resizes their storage, so the eagerly created oneMKL matrix handle (which pointed at the storage and required a queue synchronize per construction) is now created on the first oneMKL operation (oneMKL.sparse_matrix_handle) and dropped when the storage is replaced (copyto!, unsafe_free!). Empty matrices never get a handle; the wrappers short-circuit instead. Any element type can be stored; oneMKL operations still require Float32/Float64/ComplexF32/ComplexF64 with Int32/Int64 indices and throw a clear error otherwise. CSC matrices on oneMKL < 2025.3 now error at the first oneMKL operation instead of at construction.
• On-device conversions (lib/mkl/sparse_conversions.jl): CSR↔CSC↔COO, _sptranspose/_spadjoint, sparse +/- (needed by GPUArrays' issymmetric, since oneMKL has no geam-like routine), and the COO operations GPUArrays forwards to (triu, tril, kron, reshape, droptol!), built from sortperm on integer keys and broadcasts.
• Adapt/KernelAbstractions glue: device-side counterparts for kernels, get_backend, and adapt(oneArray, ::SparseMatrixCSC) returning a oneSparseMatrixCSC (as CUDA.jl and AMDGPU.jl do). This is a behavior change: previously the sparse matrix was densified.
• Docs and tests updated; test/onemkl.jl gains interface, conversion, broadcast/reduction and lazy-handle tests.

Notes

• No sparse vector type is added (oneMKL has no sparse-vector routines wrapped here).
• GPUArrays PR Generic GPU sparse arrays: storage-parametric formats owned by GPUArrays GPUArrays.jl#745 (planned 12.0) reworks the sparse backend interface; this PR targets the current 11.5 interface and will need a follow-up when that lands.

[codecov]

Codecov Report

❌ Patch coverage is 83.96501% with 55 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.76%. Comparing base (f110df6) to head (d24babb).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
lib/mkl/sparse_conversions.jl	62.96%	40 Missing ⚠️
lib/mkl/wrappers_sparse.jl	94.06%	7 Missing ⚠️
lib/mkl/array.jl	94.73%	6 Missing ⚠️
src/oneAPIKernels.jl	33.33%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #642      +/-   ##
==========================================
- Coverage   81.61%   80.76%   -0.85%     
==========================================
  Files          50       55       +5     
  Lines        3611     4087     +476     
==========================================
+ Hits         2947     3301     +354     
- Misses        664      786     +122     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.