[RNE Rewrite] Add LFM2.5 ColBERT (late-interaction multi-vector) embedding pipeline

[msluszniak]

Summary

Add a LFM2.5-ColBERT-350M late-interaction (multi-vector) embedding pipeline under the new flow. Follow-up to the single-vector embeddings work in #1292 / #1247.

Why it isn't covered by the text embedder

ColBERT is a late-interaction model: forward returns a per-token matrix [numTokens, 128] (Linear 1024→128 + per-token L2-norm), not a single pooled vector. Relevance is scored with MaxSim (for each query token, the max cosine against all document tokens, summed) rather than a single cosine. So createTextEmbedder / useTextEmbedder cannot represent it — it needs its own pipeline.

The export scripts already exist (exporter_models/lfm2_5_colbert_350m, xnnpack + mlx) and an HF repo (react-native-executorch-lfm2.5-colbert-350m, currently only v0.9.0 / old flow).

Scope

Export (export-scripts, lfm2_5_colbert_350m)

• Port both backends to the new flow (dynamic seq + get_dynamic_dims_forward companion), matching the embedding export
• MLX must be built with the software-mansion-labs/executorch fork + MLX-native quantization — a pip/torchao-built MLX delegate fails at execute with Error::InvalidProgram (same root cause as the [RNE Rewrite] Add text and image embeddings pipelines #1292 MLX fix; the MLX backend is version-locked to the runtime MLXBackend.xcframework)
• Publish xnnpack + mlx to HF react-native-executorch-lfm2.5-colbert-350m v0.10.0

Library

• New task createColbertEmbedder + useColbertEmbedder hook returning per-token embeddings [numTokens, 128]
• MaxSim late-interaction scoring helper (query vs. document token matrices)
• models.ts entries (xnnpack 8da4w + mlx int4) + index exports; download-libs already enables the MLX backend for textEmbeddings

Demo (nlp)

• Late-interaction retrieval screen (or extend Text Embeddings) using MaxSim; MLX chip disabled on Android (MLX is iOS-only)

Notes

• The exporter NOTES document that XNNPACK lowering bakes the dynamic seq dim to static [1,512] / [1,512,128] on the current ET build; decide between static + runtime padding vs. a build that preserves the dynamic seq (the embedding MLX variant shipped dynamic; both static+pad and dynamic were verified on device).
• Related: standalone on-device vector store [RNE Rewrite] Add a small standalone on-device vector store #1303.