Fix parameter forwarding in fit_predict and correct progress loss normalisation

[harens]

This PR fixes two issues in inference configuration and training progress reporting.

1. Forward fit_predict parameters correctly

Interpreter.fit_predict() previously ignored caller-provided iterations, batch_size, and verbose, always using hardcoded defaults when calling predict().

This PR forwards those arguments through so runtime behaviour matches the caller’s configuration.

---

2. Fix progress loss normalisation

Progress loss was previously accumulated as:

total_loss += loss.item() / X_.shape[1]
total_items += X_.shape[0]

This mixes units: the loss is scaled per timestep (X_.shape[1]), but averaged per sequence (X_.shape[0]), which underestimates the true mean loss.

The update instead:

total_loss  += loss.item()
total_items += X_.shape[0] * y_.shape[1]

i.e.

• accumulate the total loss over all predictions
• divide by the total number of target elements (batch_size × sequence_length)

This makes the reported loss a true per-prediction average.

This change affects only the progress display, not optimisation or model outputs.