Attach to tensor variables a member `F` that points to ndarray or symbol

This is a suggestion for a small usability improvement in the gluon API. The proposal is to include a member `F` with each tensor variable that points to mxnet `ndarray` or `symbol` module, depending on whether the variable is an NDArray or a Symbol. So for instance

x = mx.nd.zeros(10)
x.F  # this is mx.ndarray

While this is only a small change, it simplifies the API for functions and classes that do not contain learnable parameters. Such functions then don't need an explicit F parameter in the API and don't need to be wrapped in HybridBlocks.

As an example one could implement a Gaussian distribution like this:

class Gaussian:
    def __init__(self, mu, sigma):
        self.mu = mu
        self.sigma = sigma
        self.F = mu.F

    def sample(self):
        return self.F.sample_normal(mu=self.mu, sigma=self.sigma)

    def log_prob(self, x):
        return -0.5 * F.square(((x - self.mu) / self.sigma)) - F.log(self.sigma) - 0.5 * math.log(2 * math.pi)

which gives a clean API that works for both symbols and ndarrays

distr = Gaussian(mx.nd.zeros(10), mx.nd.ones(10))
s = distr.sample()
lp = distr.log_prob(s)

While one can currently shim this by using a function

def get_F(var):
    if isinstance(var, mx.nd.NDArray):
        return mx.nd
    else:
        return mx.sym

it makes sense to have a standard mechanism for this.