…on to scalar in Normal distribution

- Replace manual normal sampling and logprob calculation with utility function
- Fix tensor operations using proper broadcasting and element-wise methods
- Add ToElement import for tensor casting operations
- Improve numerical stability in probability calculations
- Generalize forward pass to work with any tensor dimension

- Remove AutodiffBackend constraint from QNet to support inference
- Extract temporal difference calculation into separate function
- Add data import for batch processing
- Clone networks for target initialization instead of loading records
- Begin implementing QNet training loop with TD targets

- Modify `train_net` to accept `Tensor<B, 2>` and `DataBatch<B>`
- Update both Q-networks with temporal difference target during training
- Add policy network entropy calculation to training loop
- Replace transition struct fields with batch-compatible states and actions

- Add Q-network forward passes for both critics
- Compute minimum Q-value between two critics
- Calculate policy loss using min Q-value and entropy
- Remove placeholder comment for optimizer (TODO remains)