Some small fixes for spike and slab Gibbs sampling.

These mainly effect cases with not much data, by addressing how the priors are set:
- the prior is set more carefully to account for missing data,
- targets for spike and slab samplers are zero'd out where they are missing,
- an optional and nonstandard update to the weight posterior precision is added, to support recreating a sampler from the literature.

PiperOrigin-RevId: 452351044

Author: ColCarroll

tensorflow_probability/python/experimental/sts_gibbs/BUILD

@@ -135,6 +135,7 @@ multi_substrate_py_library(
         "//tensorflow_probability/python/distributions:joint_distribution_auto_batched",
         "//tensorflow_probability/python/distributions:sample",
         "//tensorflow_probability/python/experimental/distributions:mvn_precision_factor_linop",
+        "//tensorflow_probability/python/internal:broadcast_util",
         "//tensorflow_probability/python/internal:parameter_properties",
         "//tensorflow_probability/python/internal:prefer_static",
         "//tensorflow_probability/python/internal:samplers",