RL: treat stop-loss and late close as pure risk management

RL: stricter PPO + looser default stop-loss for stable training

RL: tune hyperparams for stable learning with stop-loss

add hold action when spread_pct > 5% in run.py, to prevent opening new positions when the spread is too large

add stop loss feature in run.py

add the pending paper open delay to simulation the live fill delay by _update_position_from_api in live trading

add the state.position_pnl (which is the unrealized PnL) calculation in the console print

- open position is forbidden within 1 minute to close

add entry spread_pct print in run.py after the final stats

add action distribution print in run.py after the final stats

decrease the learning rate for the actor and the critic in rl_mlx.py and rl_torch.py

stage 2: add the blended reward for RL training

1. fix the _pending_orders issue, from token id level to condition id level, so for each market can only have one side (UP or DOWN) at a time

This is the version (checkpoint) that can train the model successfully.

[GOD mode training]