Refactor reset mechanisms SpikingNeuron base class and (R)Leaky definitions

[DoubleGio]

The reset mechanisms in the Leaky class got broken in version 0.08 where they got refactored a bit.
Both soft and hard reset mechanisms do not function as expected: currently the default soft reset resets to zero, whereas the hard reset resets by subtraction -- see this open issue, or you may see for yourself by running tutorial 3 with the current release of snntorch (0.9.4).
To address this:

1. Refactored SpikingNeuron parent class to more easily allow for custom reset implementations. Instead of only accepting 'subtract', 'zero', and 'none', custom reset_dict can now be redefined in child classes.
2. Fixed @reset_mechanism.setter to actually change the reset mechanism when altered. This comes with the introduction of the _set_reset_function() function, to be overwritten in child neuron classes.
3. Refactored reset mechanisms in Leaky (and RLeaky) for clarity and to function as expected:

• subtract_beta -> $U[t+1] = \beta U[t] + I[t+1] - \beta R U_{\rm thr}$
• subtract -> $U[t+1] = \beta U[t] + I[t+1] - R U_{\rm thr}$
• zero -> $U[t+1] = \beta U[t] + I[t+1] - R (\beta U[t] + I[t+1])$
• none -> $U[t+1] = \beta U[t] + I[t+1]$