Add error handling for errors in the JSON response from HuggingFaceInferenceAPIEmbeddings embed_documents call

[brian-ogrady]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

When using a custom HuggingFace API Endpoint there can be an error in the JSON response. I propose added two checks when calling the embed_documents method of the HuggingFaceInferenceAPIEmbeddings class:

1. Add a check for a JSONDecodeError
2. Add a check within the JSON response for an error

Motivation

While working with the HuggingFaceInferenceAPIEmbeddings class, I sometimes receive a nebulous error ValueError: could not convert string to float: 'e'

This results because the embed_documents method, which is called during the execution of the add_documents method of a vector store, can return a JSON response like the following: {'error': 'Input validation error: inputs cannot be empty', 'error_type': 'Validation'}

Proposal (If applicable)

The initial ValueError raised from add_documents is misleading because it makes the underlying error returned in the JSON response. I propose surfacing this error by adding a check during embed_documents to return the JSON error in a Pythonic manner.