fix(mlx): rebuild delegate with SWM executorch fork (runtime-compatible); dynamic seq + MLX-native int4

Re-export LFM2.5-Embedding to new dynamic-shape flow (BOS-prepend, no-pad, get_dynamic_dims_forward)

Annotate dynamic token dim with min/max (2..512) in config shapes

Update README.md

Add LFM2.5-Embedding-350M XNNPACK (8da4w) + MLX (int4) for react-native-executorch

initial commit