Fix training - #33

* increase by 5 times ppo continuous entropy coef

* lower continuous entropy coef for ppo

… torch error

* higher aim angle scale for potential reward
* fix default agent algorithm at CLI
* no curriculum learning : same spawn zone sizes during training

* train critic independently of kl stop for actor
* update alpha once per train call (so higher learning rate)
* higher target entropy sigma for continuous actions (SAC & PPO)

* fix distance to closest enemy at shell trajectory
* fix shell disambiguation (on pointer, not on fire pos)
* smaller shell reserve (from 30 to 10)

* re-add entropy bonus for PPO
* change fire proba target from 0.05 to 0.3

… => CLI args

* fix PPO hyperparams

* clamped alpha for ppo
* higher learning rate for ppo's alphas

* increase target_sigma (PPO & SAC)