force same gpu

[mike-ferguson]

Fix multi-GPU device mismatch in HuggingfaceSubject

Problem

When running with device_map='auto' on multi-GPU machines, inputs were always sent to cuda:0 while the embedding layer could be on another GPU (e.g. cuda:3), causing:

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:3 and cuda:0!

Small models (e.g. DistilGPT-2) were most affected because device_map='auto' can place the embedding layer on a non-zero device.

Solution

1. Use the embedding layer's device for inputs instead of assuming cuda:0: self.device = self.basemodel.get_input_embeddings().weight.device
2. In estimate_reading_times, move actual_tokens to predicted_logits.device before F.cross_entropy, since logits can reside on a different GPU with device_map='auto'.

Impact

• Fixes multi-GPU runs for all model sizes when device_map='auto' is used.
• No behavior change for single-GPU use.
• Batch jobs typically use one GPU per job, so they were unaffected.