Planning: FIP data loader

I'm starting to fit the disrnn models with FIP data inputs. This will require either a new data loader, or modification of the mice data loader. This is a place to record some planning notes

### Requirements
I need to specify the following information in the hydra model

For each input feature:
- Which fiber (0-3)
- Which channel (G, R, Iso)
- Which processing method (many possible)
- What time window (eg, 0-2s) around what time event (eg, goCue). This may involve prev/current choice indexing.

### First pass implementation
I added a new data loader that builds the FIP dataframe, z-scores the channels, and gets the average signal window

```python
fiber = 0
channel = 'G'
method='dff-bright_mc-iso-IRLS'
full_channel_name = f'{channel}_{fiber}_{method}'

nwb.df_fip = nu.create_df_fip(nwb,verbose=False)
nwb.df_fip = ed.zscore_fip(nwb.df_fip)
nwb.df_trials = tm.get_average_signal_window(
nwb,
alignment_event='goCue_start_time_in_session',
offsets = [0,2],
channel=full_channel_name,
data_column = 'data_z',
output_col='NE_FIP'
)
```

### Details to work out
- How to deal with NaNs, or missing values? It looks like one session has a single NaN. Its not clear where this is coming from. For now, I'm just filling it in with a 0 (z-scored data). But we should have a more robust checking and filling method