core/nn/modalities/text_transformer library
Text Transformer — token-index encoder.
Input: [seqLen] (1D tensor of token indices as floats — matches
the convention used everywhere else in this repo since we
don't yet have an int tensor dtype).
Output: [seqLen, embedDim] — per-position contextualized features.
Recipe: `Embedding` + `LearnedPositionalEmbedding` +
`TransformerEncoder` (no causal mask — this is a bidirectional
encoder, not a decoder). Use `poolMean` to collapse to a single
[1, embedDim] sentence embedding.