[Cleanup] Small fixes to the multi-modal models

[bohnstingl]

Description

Refine the recent Pixtral/Ministral multimodal enablement without expanding the supported model surface:

• Route all CPU↔Spyre transfers through convert(), including layout-aware transfers.
• Prepare tiled convolution weights once after model loading instead of lazily in the forward path.
• Consolidate repeated tensor-tree conversion logic in _SpyreModelWrapper.
• Reuse the shared conversion helper for model input, multimodal input, embedding-merge, logits, and model-output boundaries.
• Preserve the existing native Mistral-format Pixtral behavior and linear fast path.

Related Issues

Relates to #207

Test Plan

This PR does not introduce any functional change.

cc @nikheal2

[github-actions]

👋 Hi! Thank you for contributing.
Just a reminder: Make sure that your code passes all the linting checks, otherwise your PR won't be able to be merged. To do so, run ./format.sh.
Now you are good to go 🚀.

We also recommend installing prek and configuring it to check your code before every local commit.