Add mixed-attention Core ML mask support for stateful generation

[Skyline-23]

What

Add support for stateful Core ML language models that require multiple attention masks during generation.

Why

The current runtime only handles attentionMask / causalMask, which is not sufficient for mixed-attention Core ML exports that need separate masks for different layer types.

This change allows the stateful generation path to populate:

• fullAttentionMask
• slidingAttentionMask

when those inputs are present in the Core ML model description.

Implementation

• add fullAttentionMask and slidingAttentionMask keys to LanguageModel.Keys
• detect those inputs from modelDescription
• build additive full-attention and sliding-window masks in the stateful generation path
• resolve the sliding window size from Core ML metadata first, and fall back to Hugging Face config when needed
• factor stateful generation input assembly into a reusable helper for test coverage
• keep existing single-mask models working unchanged

Tests

• add regression tests for additive full-attention mask construction
• add regression tests for sliding-window mask construction
• add an integration-style test that verifies the full input dictionary for a mixed-attention model contract
• verify with:
  swift test --filter LanguageModelCoreMLMaskTests

Scope clarification

This PR is intended to support explicit multi-mask Core ML generation contracts in the runtime.

It does not attempt to fix exporter-side approaches that reconstruct multiple masks inside a Core ML graph from a single causalMask input.

Additional context

Closes #330

Example converted Core ML repo using the explicit multi-mask contract:
https://huggingface.co/Skyline23/translategemma-4b-it-coreml

[pcuenca]

Very interesting and cool PR @Skyline-23! I won't be able to properly test and review it until the end of the week. Meanwhile, a couple of questions:

• The converted example model seems to be using float32 instead of float16 (because of this line, and because the repo takes ~16 GB). Did you try to convert to float16? Did you try any quantization options?
• Are you using or planning to use this Core ML model in a downstream app?

Thanks a lot for the contribution!

[Skyline-23]

@pcuenca Sorry for late reply! It's fine. Please review slowly
The convert script was not optimized for the CoreML, I fixed it and add options with 4-bit and 8-bit quantized models.
It was problem with float 32 attention.
I will planning to use with on device translate app, In the test bed.
It could be canceled, but In my opinion this PR going to move repo good way for the various model support
Thanks!