[Feat]: speculative decoding support

[x-0D]

Description
Add UI and backend features to enable speculative decoding support.

ggml-org/llama.cpp#10455

Use Case
Faster interference of slower models.

https://arxiv.org/abs/2211.17192

[coder543]

I'm especially curious to see what this could do for the performance of Qwen3 4B (or maybe 8B, depending on the phone) with a Qwen3 0.6B draft model.