Atomic min max GS operator

[tuananhdao]

Changes

• Supports GS_OP_MIN, GS_OP_MAX for gs op using CUDA and HIP
• Add unit tests for the added operators, in tests/unit/gather_scatter
• Add benchmark tests to tests/bench/gs

Benchmark results

By running tests/bench/gs.

HIP, rocm/6.2.0

 -------------Mesh-------------
 Reading a binary Neko file box.nmsh
 gdim = 3, nelements =     27000
 Reading elements
 Reading BC/zone data
 Reading deformation data
 Mesh read, setting up connectivity
 Done setting up mesh and connectivity
 Mesh and connectivity setup (excluding read) time (s): 93.528476

 --------Gather-Scatter--------
 Comm         :   Device MPI
 Avg. internal:      2646000
 Avg. external:            0
 Backend      :          hip
 
GS_OP_ADD mean:   0.4491E-03, stddev:   0.3858E-05
 
GS_OP_MIN mean:   0.4492E-03, stddev:   0.3962E-05
 
GS_OP_MAX mean:   0.4490E-03, stddev:   0.4021E-05

CUDA

 -------------Mesh-------------
 Reading a binary Neko file box.nmsh
 gdim = 3, nelements =     27000
 Reading elements
 Reading BC/zone data
 Reading deformation data
 Mesh read, setting up connectivity
 Done setting up mesh and connectivity
 Mesh and connectivity setup (excluding read) time (s): 69.145985

 --------Gather-Scatter--------
 Comm         :          MPI
 Avg. internal:      2646000
 Avg. external:            0
 Backend      :         cuda

GS_OP_ADD mean:   0.7639E-04, stddev:   0.7434E-05

GS_OP_MIN mean:   0.7630E-04, stddev:   0.7027E-05

GS_OP_MAX mean:   0.7629E-04, stddev:   0.6781E-05

[MartinKarp]

Champ!