feat: Adaptive Non-uniform Timestep Sampling (arXiv:2411.09998)

Implements the paper's method for adaptively sampling timesteps during
diffusion model training by learning which timesteps yield the greatest
reduction in the objective.

Core module (library/adaptive_timestep_sampler.py):
- TimestepSamplerNetwork: MLP that outputs Beta(a,b) parameters from latents
- AdaptiveTimestepManager: Orchestrates Algorithm 1 (training) and
  Algorithm 2 (Delta approximation via queue + F-statistic feature selection)
- REINFORCE policy gradient update with entropy regularization
- Full-batch delta computation at |S| selected timesteps per paper line 7

Integration (train_network.py):
- Initialize sampler when --adaptive_timestep_sampling is enabled
- Use Beta-distributed timesteps instead of uniform random
- Algorithm 2 hooks before/after optimizer step
- Full batch delta computation via cache_batch_losses_at_S()

CLI arguments (train_util.py):
--adaptive_timestep_sampling, --adaptive_sampler_lr,
--adaptive_sampler_entropy_coeff, --adaptive_sampler_update_freq (f_S=40),
--adaptive_sampler_queue_size (|Q|=20), --adaptive_sampler_num_selected (|S|=3),
--adaptive_sampler_hidden_channels (128), --adaptive_sampler_hidden_depth (2)

Tests: 25 tests covering sampler network, F-statistic feature selection,
manager lifecycle, delta approximation, policy gradient update, cached t
REINFORCE fix, and full batch extension.

Author: 67372a