Fix conversion of some BERT embedding models

[christianazinn]

BERT-based embedding models require the use of convert-hf-to-gguf.py to be converted from safetensors/PyTorch format to GGML. The BertModel class was missing logic that resolves unsupported datatypes, resulting in some models like acge_text_embedding failing with TypeError: Got unsupported ScalarType BFloat16 when running the line data = data_torch.squeeze().numpy().

This is rectified by converting unsupported datatypes to f32 - this is done in every other model class, so it was probably just missed in BERT models. This fix allows for satisfactory conversion and subsequent quantization, as seen in this GGUF quantization of acge_text_embedding created with this fix.

It's also literally just two lines of code, so unless converting tensors in BERT models specifically is undesirable, I hope this is an easy fix.

[christianazinn]

Fixed whitespace issues, would like a review.

[iamlemec]

Makes sense to me. Can confirm that conversion of acge_text_embedding works and model outputs accurate results. Also checked that bge-base-en-v1.5 and nomic-embed-text-v1.5 give identical coverted GGUFs between this PR and master.