Fixed DEMove Implementation

[StarostinV]

Problem:

• I observed that the current implementation of DEMove, although stated to follow Nelson et al. 2013, does not actually align with the paper's approach. The existing implementation lacked affine-invariance, which is fixed in Nelson's paper.
• There was potential to optimize performance via vectorization.

Changes

Changes are only made in the implementation of DEMove in de.py.

• Alignment with Nelson et al. 2013:
  Fixed the discrepancy between the stated and actual approaches. The normal noise $\epsilon \sim N(0, \sigma^2)$ used to be added to the proposal as per $\gamma_0 (x_i - x_j) + \epsilon$. In this case, $\sigma$ should be tuned for each dimension, which confronts the idea of affine-invariance. Now it defines the distribution over gamma: $\gamma = \gamma_0 (1 + \epsilon)$, which follows Nelson et al. 2013 and ensures affine-invariance.

• Vectorization for improved performance:
  Removed a loop in favor of vectorized operations. Benchmarks showed a speed boost ranging from 5-10 times depending on the number of walkers.

• Simplified walker splitting:
  The new implementation requires splitting the walker population into two sets as opposed to the previous three-set approach.

[dfm]

Thanks! This looks good to me!