Add hipSPARSE tests (csrgemm, sptrsv, sptrsm, csr2csc, csr2coo)

[Ozone-Whiskey]

Summary

Seeds hipSPARSE test coverage (previously zero hipSPARSE tests existed) for the roc/hipSPARSE top-routine list, recycling the merged rocSPARSE tests. s/d/c/z where applicable, in both the Fortran 2003 and Fortran 2008 binding tiers.

Routines covered (17 total, f2003 + f2008 each = 34 new test programs):

Family	Routines
Sparse GEMM	Scsrgemm Dcsrgemm Ccsrgemm Zcsrgemm
Triangular solve (vector)	Ssptrsv Dsptrsv Csptrsv Zsptrsv (generic SpSV)
Triangular solve (matrix)	Ssptrsm Dsptrsm Csptrsm Zsptrsm (generic SpSM)
CSR → CSC (transpose)	Scsr2csc Dcsr2csc Ccsr2csc Zcsr2csc
CSR → COO	Xcsr2coo (index-only)

Each routine is one commit (f2003 + f2008 pair + CMake registration).

File extensions (.f08)

Per review feedback on #466/#468, files in test/f2008/ use the .f08 extension so CMake selects the Fortran 2008 standard. The shared hipfort_add_test helper is updated to prefer .f08 and fall back to .f03, so existing .f03 f2008 tests keep building.

Verification approach

Every test is self-verifying against a fixed small CSR matrix:

• csrgemm: B = A, compute C = A*A, check structure + values (complex uses a dense matmul reference).
• sptrsv / sptrsm: pick a known solution, form the RHS from a dense product, solve, confirm the solution is recovered.
• csr2csc: convert and check against the known transpose (values permuted, not conjugated).
• csr2coo: check the expanded COO row indices.

SpMV / SpMM intentionally excluded (binding bug)

SpMV/SpMM are not included. hipfort's generated hipfort_hipsparse_enums.F90 assigns wrong values to the hipsparseSpMVAlg_t / hipsparseSpMMAlg_t algorithm enums: the C headers define these as a single enum with duplicate values (the SPMV_* / SPMM_* families restart at 0, reusing the legacy MV_* / MM_* values), which a Fortran enum, bind(c) cannot represent — so the generator numbered them sequentially. For example the C HIPSPARSE_SPMV_ALG_DEFAULT = 0 but hipfort emits 4; passing it selects a COO algorithm on a CSR matrix and returns HIPSPARSE_STATUS_INVALID_VALUE. The correct fix belongs in the generator (emit these as integer parameters with the real values, as it already does for HIPSPARSE_SPMV_CSR_ALG3). rocSPARSE already covers SpMV/SpMM, so no coverage is lost here.

Notes

• All descriptor/array arguments in the generic API are type(c_ptr)-only (no array overloads), so the f2008 tests pass device pointers via c_loc(...), matching the rocSPARSE generic tests.
• All 34 new tests pass on gfx90a (MI210), ROCm 7.1.0, gfortran 13.3, both tiers. f2003 tests also compile with the Fortran 2008 pointer-interface macro OFF.

Test plan

• Both tiers build and PASS for all 17 routines on gfx90a
• f2003 tests compile with the Fortran 2008 pointer-interface macro OFF
• f2008 files use the .f08 extension

[Ozone-Whiskey]

Resolved merge conflicts

[amontoison]

LGTM 👍
Did you forget to push the tests for csr2coo?

[Ozone-Whiskey]

PR is up for csr2coo. There was a conflict with new bindings generated in develop and it before.