fix(vulkan): refresh ggml for AMD submission handling

[LauraGPT]

Summary

• refresh the pinned llama.cpp revision to 803b7fca, including AMD submission-size heuristics, the batching-threshold fix, and DeviceLost diagnostics
• adapt Fun-ASR-Nano to the updated penalties sampler API
• document Windows AMD Vulkan diagnostics and the reliable CPU fallback

Refs #3479.

Verification

• clean CPU configure and build: llama-funasr-cli, llama-funasr-paraformer, llama-funasr-sensevoice
• clean Linux Vulkan configure and build: llama-funasr-sensevoice
• confirmed linkage to libvulkan.so.1
• verified the pin contains llama.cpp commits 1051ecd2, 51eae8cf, 0bbc87b1, and 803b7fca
• git diff --check

The Windows AMD runtime result still needs reporter hardware validation; this PR intentionally does not claim the driver crash is fixed.