fix: encode() and decode() read vocab arrays in wrong direction

[CodeWithKyrian]

This PR fixes a bug where FallbackModel::encode() and FallbackModel::decode() read from the wrong internal arrays, causing every lookup to miss and fall back to the unknown-token placeholder.

Motivation and context

The FallbackModel stores two vocabularies: a forward mapping from tokens to IDs, and a reverse mapping from IDs back to tokens. The class had these two arrays stored under variable names that suggested the opposite direction — so encode() was reading from the reverse map and decode() was reading from the forward map. For CTC tokenizers like Wav2Vec2 where token strings are purely alphabetic characters (e.g. 'e' → 5) and IDs are integers, both lookups silently returned nothing and fell back to <pad> or an empty string. This produced completely silent ASR output for every Wav2Vec2 pipeline.

What's changed

• Fixed encode() to read from the token→ID mapping instead of the reverse map
• Fixed decode() to read from the ID→token mapping instead of the forward map

Breaking changes

None.