Torchfunc

This minor release introduces Python `3.6` compatibility hence the library can now be freely used on [Google's Colab](https://colab.research.google.com/).

`torchfunc.modules` module was also extended, additions being:
- `device` function (so you can check `device` of PyTorch module or `torch.Tensor`)
- `switch_device` context manager which cast `obj` (e.g. `torch.nn.Module` or `torch.Tensor`) to specified device when with-in the block and casts it back after the block is finished
- `{weight, bias, named}_parameters` - yields parameters only if they are `{weight, bias, named}` in order to remove unnecessary `if` statements and clarify the intent.

A lot of breaking changes introduced in this release.
This one can be considered as first semi-stable with features working correctly (or seemingly correctly).

Major Features and Improvements:

- New package `hooks`, where `recorders` are now located
- New module within `recorders`, `registrators`, responsible for easier registration of hooks based on indices within network or types of it's submodules/children.
- `plot` module removed and will probably be featured in separate library
- `torchfunc.performance.tips` now parses `torchfunc.performance.technology.TensorCores` tips
- Each object within `torchfunc.performance` package have now `tips()` method returning `str` describing steps one can take in order to possibly improve specific `torch.nn.Module` performance.
- Highly improved test coverage

Bug fixes:
- Proper registration of hooks via `indices` and `types`
- Fix object's representation (previously overriden by `dataclasses.dataclass`)

Hello :smile:,

This is initial release of `torchfunc` library, which currently should be considered as __alpha__.

To see what it's all about, check [README.md](https://github.com/szymonmaszke/torchfunc/blob/master/README.md).

To get in-depth info, check [documentation](https://szymonmaszke.github.io/torchfunc/).

__Just hoping this will help you with day-to-day neural net tasks as it helped me__ :100: