Add opt-in encoder-decoder framework modules

[0xShug0]

Summary

Extract the reusable encoder-decoder framework additions onto current main, without bringing in model implementations or registrations.

• Add depthwise convolutional subsampling with optional masks at each downsampling stage.
• Add a shared channel-affine weight type and BatchNorm evaluation weight binding.
• Add opt-in contiguous GLU gate materialization, propagated through Conformer modules.
• Add a ReLU feed-forward option and a cached Transformer decoder block using caller-owned KV storage and packed projections.
• Add cached flash cross-attention with an explicit additive F16 mask, exposed through a default-off decoder-block option.

Compatibility and Scope

Existing defaults remain unchanged: GELU activation, non-materialized GLU gates, and the existing cross-attention path. Existing keep-mask, attention-prior, and attention-output APIs are retained; the flash path is a separate explicit API.

Includes focused framework unit coverage for defaults, opt-ins, stage masks, affine binding, cached decoder composition, and flash-mask contracts.

Only framework headers/sources and their unit tests are included. No model code, model specs, UI, weights, build registration, or ggml changes. The Qwen prefill/cache additions already merged into main are excluded.