Preserve input dtype in population code helpers

[rlogger]

Fixes #429, and the ce_count_loss float64 case from #428.

_population_code in both snntorch/functional/acc.py and snntorch/functional/loss.py allocates its accumulator with torch.zeros(...), which always produces float32 regardless of the input dtype. This causes two problems with float64 inputs:

1. Silent precision loss in accuracy_rate: close float64 scores are downcast to float32, which can collapse them to equal values and flip the predicted class (the reproduction in snnTorch population-coded accuracy can mis-score close float64 predictions #429 reports 0.0 accuracy when the target class has the strictly larger score).
2. RuntimeError in ce_count_loss(population_code=True): the float32 population code is passed to NLLLoss alongside float64 class weights, raising expected scalar type Float but found Double.

The fix allocates the accumulator with dtype=spk_out.dtype (and passes device= directly instead of the extra .to(device) copy) so the population code always matches the input.

Added regression tests for both cases:

• test_accuracy_rate_population_code_dtype: checks that close float64 scores keep their ordering and the accuracy is scored correctly, plus a couple of basic accuracy_rate tests since acc.py had no coverage before.
• test_ce_count_loss_population_code_dtype: checks that population-coded ce_count_loss works with float64 inputs/weights and returns a float64 loss.

Full test suite passes locally (150 passed + NIR tests), and the changed files are formatted with black. float32 behavior is unchanged.

Made with Cursor