Initial pull request

[JanJereczek]

The present PR includes most of the basic functionalities for TransitionIdentifiers.jl, including:

• Windowing
• Smoothing and the underlying kernels
• De-trending
• Computation of moments of time series (mean, variance, skewness, kurtosis)
• Computation of AR1 regression model with white noise assumption
• Computation of low-frequency power spectrum
• Trend metrics via ridge regression or Kendall-tau
• Sliding window application of the above-mentioned functionalities
• Fourier surrogates (in future possibly use TimeseriesSurrogates.jl but for now reduce dependencies)
• Percentile significance test of the original time series compared to its surrogates.
• GPU acceleration for the most time consuming steps. Does not require extra syntax, just conversion of input to CuArray.

An overview of this is given in the following presentation (accessible by anyone with link): https://docs.google.com/presentation/d/1YU-hn-3Hy_p2f3tWdFFx0MWVAXVFy6Xg1lF78U8FlmM/edit?usp=sharing