Batched NNLS - #8402

GPU batch-friendly implementation of Lawson and Hanson NNLS algorithm.

Solves one problem per block: $\arg \min_x {||Ax - b ||}_2^2$ subject to $x \ge 0$ where $A \in \mathbb{R}^{n \times m}, b \in \mathbb{R}^{n}$. The batching happens by building a matrix $B^{n \times p}$ from columns $b$ and masking some columns of $A$ for each problem in a batch with a bitmap of size $m \times p$.

The current version keeps $O(n)$ data in shared memory, which limits the scaling over n.