feat(embed): add support for audio inputs

[Nina-mvH]

Adds support for embedding audio inputs

This change required adding the helper load_audio; this is a duplicate of the helper created in #176 and should be removed once that helper is available in the shared utils.

Tested with:

• facebook/wav2vec2-base
• openai/whisper-tiny
• laion/clap-htsat-unfused
  All these models produced some error messages/warnings at the beginning of the log (example below from the whisper log), but otherwise worked seemingly fine (produced an embedding vector)

Loading weights: 0it [00:00, ?it/s]
[transformers] �[1mWhisperEncoder LOAD REPORT�[0m from: openai/whisper-tiny
Key                                                           | Status     | 
--------------------------------------------------------------+------------+-
model.encoder.layers.{0, 1, 2, 3}.self_attn.out_proj.bias     | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.fc1.weight                  | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn.k_proj.weight     | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.final_layer_norm.weight     | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn.v_proj.weight     | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn.v_proj.bias       | UNEXPECTED | 
model.encoder.conv1.bias                                      | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn.out_proj.weight   | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.fc2.bias                    | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn_layer_norm.weight | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.final_layer_norm.bias       | UNEXPECTED | 
model.encoder.conv1.weight                                    | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.fc1.bias                    | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.fc2.weight                  | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn.q_proj.bias       | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn.q_proj.weight     | UNEXPECTED | 
model.encoder.conv2.bias                                      | UNEXPECTED | 
model.encoder.conv2.weight                                    | UNEXPECTED | 
model.encoder.layer_norm.weight                               | UNEXPECTED | 
model.encoder.layers.{0, 1, 2, 3}.self_attn_layer_norm.bias   | UNEXPECTED | 
model.encoder.layer_norm.bias                                 | UNEXPECTED | 
model.encoder.embed_positions.weight                          | UNEXPECTED | 
layers.{0, 1, 2, 3}.fc2.bias                                  | MISSING    | 
layers.{0, 1, 2, 3}.fc2.weight                                | MISSING    | 
layers.{0, 1, 2, 3}.self_attn.v_proj.bias                     | MISSING    | 
layers.{0, 1, 2, 3}.final_layer_norm.bias                     | MISSING    | 
layers.{0, 1, 2, 3}.self_attn.k_proj.weight                   | MISSING    | 
layers.{0, 1, 2, 3}.self_attn.v_proj.weight                   | MISSING    | 
layers.{0, 1, 2, 3}.self_attn_layer_norm.weight               | MISSING    | 
layers.{0, 1, 2, 3}.self_attn.q_proj.weight                   | MISSING    | 
layers.{0, 1, 2, 3}.final_layer_norm.weight                   | MISSING    | 
layers.{0, 1, 2, 3}.self_attn.out_proj.weight                 | MISSING    | 
conv1.bias                                                    | MISSING    | 
conv2.bias                                                    | MISSING    | 
layers.{0, 1, 2, 3}.self_attn.out_proj.bias                   | MISSING    | 
embed_positions.weight                                        | MISSING    | 
layers.{0, 1, 2, 3}.fc1.bias                                  | MISSING    | 
layers.{0, 1, 2, 3}.self_attn.q_proj.bias                     | MISSING    | 
layers.{0, 1, 2, 3}.fc1.weight                                | MISSING    | 
conv1.weight                                                  | MISSING    | 
layers.{0, 1, 2, 3}.self_attn_layer_norm.bias                 | MISSING    | 
conv2.weight                                                  | MISSING    | 
layer_norm.weight                                             | MISSING    | 
layer_norm.bias                                               | MISSING    | 

Notes:
- UNEXPECTED:	can be ignored when loading from different task/architecture; not ok if you expect identical arch.
- MISSING:	those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.
2026-07-08 16:13:16 | INFO     |    Embedding dimension: 384

Closes #244