[riktor]

The CPU CG solver uses BLAS symv for its products with the Gram matrix
YtY: once per row for the initial residual, and once per CG step for A·p.
calculate_loss also calls symv once per row. With the pthreads OpenBLAS
build used in these benchmarks,
concurrent symv calls scale poorly beyond a few OpenMP threads. On last.fm,
CG was no faster with 40 threads than with 4.

This replaces each symv call with n BLAS level-1 axpy calls, where n
is the number of factors. The matrix-vector product is mathematically unchanged;
results differ only in floating-point rounding due to the changed summation order.

Seconds per ALS iteration on a dual-socket Xeon (20 cores per socket),
128 factors, cg_steps=3, one thread per physical core,
OPENBLAS_NUM_THREADS=1, before → after:

threads	last.fm 360k	ml20m
1	14.3 → 15.1	10.0 → 11.0
8	5.1 → 2.4	2.2 → 1.5
40	6.0 → 0.7	1.9 → 0.4

Single-threaded runs are about 10% slower (medians of five runs each:
last.fm 14.3 → 15.9 s, ml20m 10.3 → 11.1 s): without thread contention,
one symv call is faster than n axpy calls.