Fix to_numpy() crash on bfloat16 tensors by upcasting to float32

[robbiebusinessacc]

transformer_lens.utilities.tensors.to_numpy (re-exported as
transformer_lens.utils.to_numpy) raises TypeError: Got unsupported ScalarType BFloat16 when passed a bfloat16 tensor, because NumPy has no
bfloat16 dtype and torch.Tensor.numpy() cannot convert it directly.

bfloat16 is common in TransformerLens since many pretrained models load in
reduced precision, and to_numpy is used in several utility paths (e.g.
utilities/slice.py), so this is easy to hit.

Fix: in the torch.Tensor / nn.Parameter branch, detach + move to CPU,
then upcast to float32 when the dtype is bfloat16 before calling
.numpy(). All other dtypes (float32, float16, int, etc.) are unchanged.

Tests: added a TestToNumpy class covering bfloat16 tensors and
nn.Parameter, float32/float16/int passthrough, numpy-array identity,
list/tuple, scalars, and the invalid-type ValueError. Reverting the fix
makes the two bfloat16 tests fail with the original TypeError, confirming
they're meaningful.

No existing tracking issue.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality
  to not work as expected)
• This change requires a documentation update

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature
  works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect
  backward compatibility