[Multivec 3/5]: Sequential SpMV updates

[victorapm]

This is part of a series of PRs for enabling BoomerAMG to be applied to multivectors.

Changes in this PR:

1. Extends hypre's SpMV on GPUs to work with multivectors.
2. Improves hypre's SpMV on CPUs when using multivectors.
3. Adds calls to cusparse's SpMM

Regression tests:

• [x] Tux
• [x] Lassen

[victorapm]

Times (in seconds) for 1000 SpMV (only one vector component) with 7-pt stencil Poisson operator:

Nx	Ny	Nz	DOFs	Nonzeros	master	multivec	cuSPARSE 11.2.0
64	64	64	262,144	1,810,432	0.041974	0.044754	0.061116
96	96	96	884,736	6,137,856	0.125632	0.129209	0.165329
128	128	128	2,097,152	14,581,760	0.288029	0.286971	0.344246
160	160	160	4,096,000	28,518,400	0.552682	0.552286	0.645487
192	192	192	7,077,888	49,324,032	0.951319	0.946958	1.091524

Times (in seconds) for 1000 SpMV (only one vector component) with 27-pt stencil Poisson operator:

Nx	Ny	Nz	DOFs	Nonzeros	master	multivec	cuSPARSE 11.2.0
64	64	64	262,144	6,859,000	0.120503	0.122454	0.135166
96	96	96	884,736	23,393,656	0.388851	0.389597	0.435317
128	128	128	2,097,152	55,742,968	0.983655	0.985481	0.9982
160	160	160	4,096,000	109,215,352	2.011068	2.021667	1.950247
192	192	192	7,077,888	189,119,224	3.502067	3.512579	3.392852

Legend:

1. master: uses master branch
2. multivec: this branch
3. Platform: one V100 (Lassen)