[Feature] Generalize parameter constraints over certain range

[jpmoutinho]

When running parameterized digital-analog programs parameters often have to be optimized while bounded over a specific range. For example, currently in the AddressingPattern class this is done by re-defining the parameters with a multiplication with a Heaviside function such that values outside the allowed range are counted as zero. To note that for optimization purposes, a smooth approximation of the Heaviside needs to be used for gradient calculation, as currently done in the addressing pattern.

This can potentially be generalized, either as a general function or directly inside the Qadence VariationalParameter to be used in other modules.

Potentially related: can parameters like angles be defined mod 2 pi?

[jpmoutinho]

This is potentially needed to deal with the domain of certain functions in #267

E.g., the trainable frequencies for a chebyshev feature map should respect the domain of acos being (-1, 1). Since there is already a target_range parameter that re-scales the data to that range, it would be enough that the trainable frequencies stay inside the (0, 1) range.

A few potentially relevant links from the web:

https://discuss.pytorch.org/t/set-constraints-on-parameters-or-layers/23620/3
https://discuss.pytorch.org/t/restrict-range-of-variable-during-gradient-descent/1933

[mlahariya]

Might be useful but low priority. Check its usefulness for Qadence 2.
@mlahariya, check its usefulness and if needs to be implemented.