Improve pickle (CandleTensor) conversions to NestedValue

[laggui]

Follow-up to PR #1751.

Reposting the comment as a new issue.

Instead of serializing CandleTensor, we should just convert the underlying ParamSerde and DataSerialize structures to NestedValue directly. The data (vector) pointer will be copied to NestedValue::F32s or NestedValue::F16s (a new datatype that preserves encoding). You will need to modify the serialize_data function at https://github.com/tracel-ai/burn/blob/main/crates/burn-import/src/pytorch/reader.rs#L127-L145.

You will need to do something similar to the serialize function in https://github.com/tracel-ai/burn/blob/main/crates/burn-import/src/pytorch/adapter.rs#L68-L78. The logic overlaps with converting ParamSerde::new(param_id, DataSerialize::new(data, shape)) to NestedValue.