GPU kernels for PBC overlap integrals (#456)

* Add overlap integral kernel

* Add ipovlp and ipkin kernels

* cleanup int1e

* Replace ovlp and kin integrals in pbc.scf and pbc.grad

* Fix pbc int1e interface

* Fix lattice sum range and other bugs

* Fix np and cp compatibility

---------

Co-authored-by: Qiming Sun <qiming.sun@bytedance.com>

Author: sunqm

gpu4pyscf/lib/pbc/CMakeLists.txt

@@ -3,6 +3,7 @@ set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --ptxas-options=-v")# -maxrregcount=12
 add_library(pbc SHARED
   pbc_driver.cu ft_ao.cu unrolled_ft_ao.cu ft_ao_bdiv.cu
   fill_int3c2e.cu unrolled_int3c2e.cu fill_int2c2e.cu
+  overlap.cu
   estimator.cu
   rys_roots_dat.cu
   nr_eval_gto.cu

gpu4pyscf/lib/pbc/fill_int2c2e.cu

@@ -57,7 +57,7 @@ void pbc_int2c2e_kernel(double *out, PBCIntEnvVars envs, PBCInt2c2eBounds bounds
     double *env = envs.env;
     double *img_coords = envs.img_coords;
 
-    int gout_stride = bounds.gout_stride_lookup[lj*L_AUX1+li];
+    int gout_stride = bounds.gout_stride_lookup[li*L_AUX1+lj];
     int nsp_per_block = THREADS / gout_stride;
     int sp_id = thread_id % nsp_per_block;
     int gout_id = thread_id / nsp_per_block;