Add seperate config for predict batch size and train batch size.

Updating model weights takes alot more GPU memory than just a forward model pass.

predict.tile is slower than it needs to be because its using trainer.predict, which inherits a dataloader with batch size set by the global config https://github.com/weecology/DeepForest/blob/3dbc8342de766f1f504d3c2da69c1fbc2443da42/src/deepforest/main.py#L348

and in train gets from load_dataset.

https://github.com/weecology/DeepForest/blob/3dbc8342de766f1f504d3c2da69c1fbc2443da42/src/deepforest/main.py#L335

the default is 1 because training is unknown size GPU (probably should be 2)

1. Make a predict_batch_size and a train_batch_size config arg
2. Update defaults to 2 for train and 8 for predict.
3. update the config doc
4. Write tests showing the dataloaders of each are yielding correct sizes.

I'm unsure about the val dataloader batch size, maybe should be higher, not clear to me the GPU memory. I think val batch size should be the predict size, since no weights are updated.

Start in src/deepforest/main.py around the dataloader setup at lines 335 and 348, then trace how the global batch-size config reaches training and prediction. Add separate train and predict settings with the requested defaults, resolve whether validation follows prediction, update the config documentation, and add tests that verify each dataloader batch size.