Persistent GradientTape hangs with Distribution.sample()

Drawing samples from an instance of tfp.distributions.Distribution in a persistent=True GradientTape context seems to hang:

import tensorflow as tf
import tensorflow_probability as tfp

alpha = tf.Variable([2., 0.5])
beta = tf.Variable([2., 3.])

with tf.GradientTape() as tape:
    sample = tf.random.gamma(shape=(1,), alpha=alpha, beta=beta)

print(sample)

with tf.GradientTape() as tape:
    sample = tfp.distributions.Gamma(concentration=alpha, rate=beta).sample()

print(sample)

with tf.GradientTape(persistent=True) as tape:
    sample = tf.random.gamma(shape=(1,), alpha=alpha, beta=beta)

print(sample)

with tf.GradientTape(persistent=True) as tape:
    sample = tfp.distributions.Gamma(concentration=alpha, rate=beta).sample()

print(sample)  # FIXME: never reachers here!!!

This seems to happen for TF in [2.3.0, 2.3.1, nightly], TFP in [0.11.0, 0.11.1, nightly], Python 3.7, Colab and Mac OS (have not tested release and platform combinations).

Start with the Python reproduction in the issue, comparing direct tf.random.gamma sampling with tfp.distributions.Gamma.sample() under persistent and non-persistent GradientTape contexts. Confirm whether the final TFP sample still hangs across the listed TensorFlow and TensorFlow Probability versions; done means identifying and resolving the persistent-tape hang without regressing the working cases.