Cplxmodule

- COSMIT: accepted [Black](https://black.readthedocs.io/en/stable/index.html) a the code style of choice, introduced pre-commit hooks for developers
- FIX: having a dunder-version in the root of the package is a the standard (issue 24)
- FIX: set the minimal python to `3.7` as pointed out in issue 24
- UPD: bumped the base version of torch to at least `1.8`
- FIX: upgraded `.utils.spectrum` to new native torch complex backend (`torch>=1.8`)
- FIX: ensured ONNX support in PR 14
- ENH: implemented modulus-based maxpooling, requested in issue 17
- FIX: made `.Cplx` instances `deepcopy`-able, fixing issue 18
- DOC: improved docs for `.nn.ModReLU` indicating the sign-deviation from the original paper proposing it (issue 22)
- DOC: added a basic TOC to the main README docs

Completed DEPRECATION cycles

- misnamed VD and misplaced ARD layers in `.nn.relevance`
- sparsity stats badly placed in `.utils.stats`
- misnamed $\ell_0$ probabilistic pruning layer in `.nn.relevance.extensions.real`, since it had nothing to do with the Automatic Relevance Determination Bayesian approach

- FIX: Fixed shape mismatch in `.nn.init.cplx_trabelsi_independent_`, which prevented it from working properly [ 11](https://github.com/ivannz/cplxmodule/issues/11)
- ENH: [Hendrik Schröter](https://github.com/Rikorose) implemented Complex Transposed Convolutions [# 8](https://github.com/ivannz/cplxmodule/pull/8), squeeze/unsqueeze methods for `Cplx` [# 7](https://github.com/ivannz/cplxmodule/pull/7), and added support for `.view` and `.view_as` methods for `Cplx` [# 6](https://github.com/ivannz/cplxmodule/pull/6)
- ENH: Introduce converters for special torch format of complex tensors (last dim is exactly 2) see [torch.fft](https://pytorch.org/docs/stable/generated/torch.fft.html#torch.fft)
- ENH: `Cplx` now also has `.size()` method, which mimics `torch.Tensor.size()`
- DOC: Improved documentation of `.nn.casting` modules

- structure of the `.nn.relevance` was simplified
- importing from `nn.relevance.ard` has been deprecated, and ARD layers have been moved
to `.real` or `.complex` depending on their type
- changed relevance layers class hierarchy in `.relevance.real` and `.relevance.complex`:
- factored out Gaussian Local Reparameterization into pure `*Gaussian` layers,
that reside in `.real.base` and `.complex.base`
- subclassed Variational Dropout layers (`*VD`) from `*Gaussian` with improper prior KL mixin
- subclassed ARD layers (`*ARD`) from Variational Dropout layers `*VD` with ARD Gaussian prior KL mixin

Prior version used different version numbering and although the layers are backwards compatible, their location within the library was much different.