`None` gradients with bijectors

Bijector cacheing causes None gradients:

with tf.GradientTape() as tape:
  d = tfd.MultivariateNormalTriL(scale_tril=tf.eye(2))
  x = d.sample()  
  tape.watch(x)
  g = tape.gradient(d.log_prob(x), x)

g is None  # True
tfp.math.value_and_gradient(d.log_prob, x)[1] is None  # True

The basic cause of this is that bijector.inverse(bijector.forward(x)) returns x from the cache, and the gradient of this will be None.

Currently, this can be worked around in the above example by breaking the cache. For example, replacing x = d.sample() with any of

x = d.sample() + 0
x = tf.identity(d.sample())
x = tf.stop_gradient(d.sample())

Start with the Python reproducer using MultivariateNormalTriL, GradientTape, and tfp.math.value_and_gradient, then inspect the bijector cache behavior described in the issue. Done means the shown gradient calls no longer return None without relying on the listed cache-breaking workarounds.