[BugFix] Fix quant file selection in download_gguf

[ricky-chaoju]

download_gguf hands its patterns to snapshot_download, which matches them with fnmatch against repo-relative paths where * spans /, then re-matches the same patterns locally with glob, where * does not. A repo that publishes each quant in its own folder therefore downloads fine and then dies with Downloaded GGUF files not found. That is most of the big MoE repos: all 363 gguf files in unsloth/Kimi-K2-Instruct-GGUF sit in subdirectories, and unsloth/DeepSeek-V3.1-GGUF, Qwen/Qwen3-235B-A22B-GGUF, and unsloth/gpt-oss-120b-GGUF are the same. Using ** with recursive=True makes the local scan see what was downloaded.

The same patterns also match the projector. For :BF16, the candidates are mmproj-BF16.gguf and gemma-3-4b-it-BF16.gguf, and the (count("-"), name) tie-break prefers the mmproj, so the vision tower comes back as the backbone. download_mmproj already resolves the projector, so I exclude it here and rank on the file name instead of the full path.

One behaviour change worth flagging: unsloth/gemma-3-4b-it-GGUF:F16 publishes no F16 backbone, only BF16. It used to resolve to the projector and load it as the model; now it raises the existing "not found" error.

test_download_gguf_subdir mocked glob.glob into returning a subdirectory hit for any matching pattern, so it passed against code that could not do it. I left it and added four tests that build a real directory tree.

[ricky-chaoju]

@Isotr0py Can you please take a look? Thanks!