Multiple implementations of functions with the same backend

[yngvem]

Continuing of the discussion in #434 (comment). We may want to allow multiple implementations of the same function. For example, we may want to make it possible to use opt-einsum to cache contraction paths without this being the default behaviour.

Here is an example with a possible API (without any accompanying implementation).

@allow_alternative_implementations
def absolute_value(x):
    return abs(x)

@absolute_value.alternative_implementation(name="sqrt_based")
def _sqrt_based_abs(x):
    return sqrt(x*x)


absolute_value(x)  # use original implementation
absolute_value.use_alternative("sqrt_based")

absolute_value(x)  # use sqrt_based

absolute_value.use_alternative(None)
absolute_value(x)  # use original implementation again

[JeanKossaifi]

Thanks @yngvem! I really like it!

Maybe we could call it @dynamically_dispatched and then have options to dispatch. We could even use a unified system to do backend dispatching, tensor algebra backend dispatching and this.