Preserve prompt padding masks in cross-attention

[zjn20030811]

Summary

• retain the tokenizer padding mask in the MoT, Hidden, and Unified prompt encoders
• centralize padding cleanup in a vectorized helper that zeroes padded embeddings without mutating encoder output
• add CPU regressions for padded-token masking, non-contiguous masks, input validation, and non-mutating behavior

Why

All three prompt paths zeroed padded embeddings and then replaced the tokenizer mask with an all-valid mask. Zero vectors still participate in attention normalization, so the amount of padding could change text conditioning. Retaining the original mask excludes padded keys from cross-attention.

Validation

• 4 CPU regression cases passed
• syntax checks passed for all changed model modules and the regression file
• whitespace validation passed