- Adding norm_to_scale_identity_weight_per_block to multiply and upda…

…te_cache methods of estimator which allows the identity_weight to be scaled differently for each block according to some kind of norm (or norm-like function) of the curvature for that block.

- Fixing minor bug that would cause some curvature blocks to use an improperly scaled damping when multiplying with power=1 and use_cached=True, for classes that have non-trivial state_dependent_scale methods.

- Adding whitespace to improve readability.

PiperOrigin-RevId: 571364604