fix(torch): preserve state dtype on reset

[Missing-Identity]

Summary

Fixes #31 by preserving the live tensor dtype when TorchModule.reset_state() regenerates a Torch-backed state value.

• Reset state tensors now retain both their device and dtype.
• Add CPU coverage for LIFTorch.to(torch.float64) followed by reset_state().

Validation

• pytest -q tests/tests_default/test_lif_torch.py — 15 passed, 1 skipped
• pytest -q tests/tests_default/test_torch_module.py — 7 passed
• black==22.3.0 --check rockpool/nn/modules/torch/torch_module.py tests/tests_default/test_lif_torch.py
• git diff --check

Scope note

The current upstream LIFTorch float64 forward path also fails without a reset because its spike-generation implementation emits float32. This PR intentionally fixes only the reset-state dtype regression reported in #31; the broader forward-path issue is separate from this minimal change.

[ssinhaleite]

Dear @Missing-Identity ,

Thank you so much for opening a PR!
Before I review and merge it, I need you to sign our Contributor's Agreement.

Please send an email to vanessa.leite [at] synsense.ai titled "Contributor's Agreement," and I will share more information with you on how we move forward.

Best,
Vanessa Leite
Senior Software Engineer @ SynSense