remove threadprivate memaloc in omp

[hhorii]

Summary

#1384 is not enough to fix all the cases.
This fix will resolve issues related to allocate temporal memory in avx.

Details and comments

The fix of #1384 is not enough to run following codes.

from qiskit import transpile
from qiskit.circuit.library import QFT
from qiskit.providers.aer import AerSimulator

circ = QFT(20)
circ.measure_all()

simulator = AerSimulator()

result = simulator.run([transpile(circ, simulator, basis_gates=['u3', 'cx'])] * 4,
                       **{'_parallel_experiments': 2,
                          '_parallel_state_update': 2,
                          'shots': 10}).result()

Because of the structure of codes, temporal memory is allocated and freed in the other OMP blocks in avx codes.
There is no guarantee to use the same set of threads to run these blocks, especially when nested omp is enabled.

This PR changes memory allocation to malloc redundant spaces. Though memory is consumed slightly more, memory is allocated and freed always-correctly.

[jakelishman]

In the OpenMP case here, how are we ensuring that there's no thread collision in the pointer? I see that double_tmp is allocated space for 2 * omp_max_num_threads complex doubles, but _load_diagonal_input seems to access this pointer at 0 offset. Should the lambda get its thread number outside of the loop, and access double_tmp at an offset of 2 * omp_get_thread_num() (or 0 if not OpenMP)?

[hhorii]

Thanks... This is my mistake. The reason why memory is allocated with thradprivate was to avoid conflict. Allocating two double or four float are necessary for each thread. double_tmp and float_tmp must be accessed via indexers.

[jakelishman]

This looks correct to me. I had wondered if it were possible to allocate double_tmp on the stack instead of on the heap to avoid the unpleasant call to malloc, but I'm guessing the issue is that omp_get_num_threads() is only known at run-time?

[hhorii]

@jakelishman Thank you so much. Results of omp_get_max_threads() and omp_get_thread_num() depend on OMP settings and thread-scheduling. I think alloca() is not necessary because double_tmp will be always on LLC.