Warn when init_noise_std and related params are ignored with tanh_normal (#661)

[SAY-5]

Fixes #661.

Problem

When using make_ppo_networks() with the default distribution_type='tanh_normal', the init_noise_std, noise_std_type, and state_dependent_std parameters are accepted without error but have no effect. The tanh_normal branch creates a plain MLP whose output fully determines the standard deviation via softplus in NormalTanhDistribution.create_dist(). Only the normal branch uses PolicyModuleWithStd where these parameters actually initialize a learnable noise parameter.

As the reporter noted, this means any hyperparameter sweep over init_noise_std under the default tanh_normal produces identical results — wasted compute with zero feedback.

Fix

Emit a warnings.warn() listing the non-default parameters when distribution_type='tanh_normal' and any of init_noise_std, noise_std_type, or state_dependent_std differ from their defaults. The warning fires once at network construction time and includes actionable guidance ("These parameters only apply to distribution_type='normal'").

This is the least-invasive option from the reporter's suggestions — no breaking changes, no API restructure, just a clear signal that the parameters are being ignored.

Example

make_ppo_networks(obs_size, act_size, init_noise_std=0.5)
# UserWarning: init_noise_std=0.5 has no effect with
# distribution_type="tanh_normal". The standard deviation is determined
# entirely by the policy network output. These parameters only apply to
# distribution_type="normal".