Parallelize sampling measurement in MPS

[Patataman]

Summary

Hello, this PR is a parallelization over the measurement sampling algorithm implemented in #1377.

Details and comments

Measurement can be parallelized at shots level. A small critical part is necessary to ensure the random numbers are generated correctly. With this, performance is better than current implementation in almost every case.

For the PR and benchmarking, parallelism can be activated or deactivated using the environmental variable PRL_PROB_MEAS=1. If you are ok with the parallelization I will just remove the condition.

Here are some results using Random Quantum Circuit (https://arxiv.org/pdf/2207.14280.pdf) for 30 qubits and different depths:

Server configuration

CPU	Intel Xeon Gold 6148
# sockets	2
# cores	20
RAM	192GB
GPU	None
OS	Ubuntu 22.04.1 LTS
Python	3.10
OpenBLAS/LAPACK	0.3.21
gcc	v11.3

For Depth 1 and 3, the main "problem" is the overhead from the parallelization, execution time was already small.

And here the specific times (in seconds)

Depth	Base	Parallel
1	0.030685282	0.042850685
3	0.03668642	0.053231239
5	0.076008368	0.073100615
10	198.6679559	55.42104983
12	1220.6618	1025.729388
15	59207.97474	39454.64986

Or checking with a different number of qubits but same depth

[merav-aharoni]

Looks good!
The graphs are a bit strange. I don't see why the depth should be a factor in the performance improvement. Also increasing the number of qubits is not a direct factor. The main factor should be the number of shots. I think a graph where you increase the number of shots on the x-axis would be best to show your improvement.