Add detection_filter to predict() method to allow for user-defined logic

[brendancol]

Hey @giswqs...I was interested in being able to have finer-grain control over the filtering of detections, especially during batch_predict.

Just an idea, but what about adding a callable detection_filter arg to predict()?

...which expects a function like:

def max_25_detections_filter_func(box, mask, logit, phrase, i) -> bool:
    return i < 25

To then be passed in kwargs down the line...

sam.batch_predict(..., detection_filter=max_25_detections_filter_func) 

I've implemented the following above, and seems to work well, but would appreciate hearing other people's perspectives on the API for something like this.

[giswqs]

LGTM.