Studio: pin Windows/macOS/CPU prebuilts to the Unsloth llama.cpp version

[oobabooga]

Today Unsloth Studio can install a different llama.cpp version depending on the platform:

Your machine	Gets its llama.cpp build from	Which version it picks
Linux + NVIDIA (CUDA)	unslothai/llama.cpp (our daily builds)	newest unsloth release
Windows	ggml-org/llama.cpp	newest ggml release
macOS	ggml-org/llama.cpp	newest ggml release
Linux, no GPU (CPU only)	ggml-org/llama.cpp	newest ggml release

Each platform picks its own newest release, and those drift apart (right now unsloth is on b9334 while ggml is on b9370), so two users on different machines can end up on different llama.cpp builds.

This pins every platform to one version: the one the latest unslothai/llama.cpp release was built from. Where each platform downloads from does not change (the middle column stays the same); only the version does, they all use that one shared version.

setup.sh / setup.ps1 resolve that version once and pass it as --llama-tag to all prebuilt installs. Explicit pins are respected, and it falls back to the old "latest" behavior if the version can't be resolved.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[gemini-code-assist]

Code Review

This pull request standardizes the llama.cpp version across platforms in both studio/setup.ps1 and studio/setup.sh. When the requested tag is set to "latest" and no explicit release tag is specified, the scripts now fetch the upstream_tag from the latest unslothai/llama.cpp release manifest to ensure consistent prebuilt downloads. Additionally, the tag resolution logic in studio/setup.sh has been simplified. There are no review comments or feedback to address for these changes.

[oobabooga]

I'm closing this PR because #6311 will make it obsolete (artifacts will not be sourced from ggml-org/llama.cpp anymore).

cc @danielhanchen