[Code scan] Respect CPU fallback before moving pretraining models to CUDA

This issue is a result of a Codex global repository scan.

## Summary
The non-DDP pretraining trainer calls self.model.cuda() unconditionally, then later checks whether CUDA is available and disables fp16. On CPU-only machines the trainer fails during construction before the fallback check can run.

## Code references
https://github.com/deepmodeling/unimol_tools/blob/4596596aa8f73eb462d5cc5a921d79966d0465da/unimol_tools/pretrain/trainer.py#L44-L67
https://github.com/deepmodeling/unimol_tools/blob/4596596aa8f73eb462d5cc5a921d79966d0465da/unimol_tools/pretrain/trainer.py#L164-L174

## Impact
Pretraining cannot run or be smoke-tested on CPU-only environments despite having later logic intended to disable fp16 when CUDA is unavailable.

## Suggested fix
Choose self.device first, disable fp16 before moving modules, and use model.to(self.device) and loss_fn.to(self.device) instead of unconditional .cuda() calls.

Start with unimol_tools/pretrain/trainer.py at the referenced lines 44-67 and 164-174, tracing device selection and the CPU fallback during trainer construction. Verify the trainer can initialize on a CPU-only machine, disables fp16 before device transfers, and places the model and loss function on the selected device; confirm this with a CPU-only pretraining smoke test.

Written by the indexing model from the issue text.