Kernel speed improvements

[rileyjmurray]

Collects sparse-SpMM kernel/benchmark work and first-class MKL sparse-BLAS support
(plus the CI plumbing to actually build and test it), Non-MKL builds and the full test suite
are unaffected throughout; the MKL sparse path is now exercised in CI for the first time.

Sparse SpMM kernels

Speed up CSR ColMajor left_spmm ~1.8–2× (5ae2093). apply_csr_to_vector_ik the hot inner kernel behind apply_csr_jik_p11) gets a #pragma omp simd reduction on the inner sparse dot-product — the win is FP reassociation (multiple partial sums break the serial accumulator chain; isolated, this is the entire effect of ffast-math on M3, while -ffp-contract=fast and -mcpu=native do nothing). A unit-stride if constexpr specialization (incv==incAv==1, the ColMajor case) drops index multiplies for another ~5–8%. Also fixes a latent int→int64_t truncation of the column index in the general-stride path. Measured (M3, density 0.01, min µs):

config	OMP=1	OMP=4
2000² d=200	3886 → 2169	1188 → 621
2000² d=2000	44197 → 22156	12730 → 6959

CSC was evaluated but its scatter kernel is latency-bound and saw no benefit, so it's left unchanged.

MKL sparse BLAS

• Handle CSC in mkl_left_spmm as a CSR-of-transpose view (7fb84cd).
  MKL's mkl_sparse_?_mm rejects CSC, so CSC previously fell back to the
  hand-rolled scatter kernel. But a CSC matrix's (colptr, rowidxs, vals) arrays
  are a CSR matrix for Aᵀ (exactly transpose_as_csr's zero-copy view), so
  we build that CSR-of-Aᵀ handle and flip the MKL operation: NoTrans → TRANSPOSE,
  Trans → NON_TRANSPOSE. Nothing is copied (MKL supports SPARSE_OPERATION_TRANSPOSE
  on CSR natively). With CSC handled, the opA=Trans branch in left_spmm no
  longer needs to pre-route a transposed CSR to MKL to dodge a CSC fallback — that
  special case is removed; the transpose-and-recurse path now reaches MKL for
  every format.
• CI: build and actually test the MKL sparse path (02ffeaa, 6e1e7a6, 1290756).
  Debian's libmkl-dev puts headers under /usr/include/mkl with no MKLROOT,
  so blaspp never defined BLAS_HAVE_MKL and RandBLAS logged
  Checking for MKL sparse BLAS ... FALSE — the entire RandBLAS_HAS_MKL path
  (spgemm + accelerated spmm, including the new CSC handling) was silently
  skipped. The setup-randblas-deps action now discovers the MKL header dir and
  exports MKLROOT/CPATH, MKL_sparse.cmake's find_path handles both oneAPI
  and Debian layouts, and the MKL job builds blaspp/lapackpp as ILP64 (so
  MKL_INT is 64-bit, matching RandBLAS's int64_t sparse indices required by
  spgemm's static_assert). ILP64 is scoped to the MKL backend only, to avoid
  breaking the LP64 OpenBLAS jobs.

SpMM benchmark — examples/simple-kernel-benchmarks/spmm_performance.cc

• Route every kernel through left_spmm using both dense layouts (568e89b).
  The old benchmark used left_spmm (ColMajor) + right_spmm (ColMajor); since
  right_spmm only reaches the RowMajor kernels by transposing the sparse
  operand, the same coverage comes from left_spmm with ColMajor and RowMajor
  dense operands (ColMajor → jik/jki, RowMajor → ikb/kib; COO delegates to
  CSC). One logical dense matrix is stored in both layouts so the two are directly
  comparable against a single densify+GEMM oracle.
• Add opB=Trans (C = A·Bᵀ) coverage (6598d08). This is the only
  opA=NoTrans path MKL declines, and it exercises the strided (incv/incAv≠1)
  branch of the CSR kernel that the unit-stride path never hits — so the kernel
  speedup above widens the measured NoTrans-vs-Trans gap. Two extra tables +
  6 correctness checks; the transposed operand reuses the existing buffers (the
  transpose of a ColMajor n×d buffer is the RowMajor buffer).
• Consistent (A, B, C) = (sparse, dense, result) notation (a857648),
  matching the left_spmm/spmm API (the file previously mixed (S, A, B)).

Misc

• Cache-efficiency improvement to potrf_upper_sequential (34c6dad) in the
  test LAPACK-like helpers.
• Fix a -Wsign-compare warning in test_exceptions.cc (0bfe539) — a
  misplaced (int) cast compared find()'s size_type to npos; swept all
  test + example sources, this was the only occurrence.

Testing

• ctest 429/429 on non-MKL builds (macOS + Linux OpenBLAS).
• The MKL sparse path — spgemm and accelerated spmm including the new CSC
  handling — is now compiled and run by the linux-gcc-mkl (ILP64) job; this is
  the first CI coverage of RandBLAS_HAS_MKL.
• CSR kernel speedup benchmarked on Apple M3 (numbers above).