Discover models across all drives; group split GGUF shards

[rjckkkkk]

On Windows the model scanner only looked at %USERPROFILE% directories, so
models on other drives (D:\models, D:\lmstudio\models, ...) were never
found and users had to set AIMA_MODEL_DIR by hand. Split GGUF shards were
each registered as a separate model, and mmproj projector files showed up
as standalone models — both break one-click deploy from the Web UI (a user
clicking a single shard would fail to load the model).

Scan discovery:

• enumerate fixed/removable drives and probe conventional model dirs that
  exist (:\models, :\lmstudio\models, .ollama, HF hub). This
  stays targeted — it never walks whole drives, so recycle bins and system
  folders are not pulled in.
• AIMA_MODEL_DIR now accepts a list (os.PathListSeparator) for custom dirs.
• honor HF_HOME / OLLAMA_MODELS for relocated caches.
• harden scanner.yaml skip list ($recycle.bin, system volume information,
  node_modules, .git, ...) so even an explicit drive-root scan stays clean.

GGUF grouping:

• collapse "...-00001-of-000NN.gguf" shards into one logical model whose
  Path is the first shard (llama.cpp auto-loads the rest) and whose size is
  the sum of all parts — so deploy targets the whole model, not one piece.
• drop mmproj projector files. The scanner already skips "mmproj" dirs
  (scanner.yaml); this covers the file-name form, so projectors no longer
  appear as deployable models.

Pure logic (groupGGUFModels, dedupePaths, DefaultScanPaths multi-path) is
table-driven tested. Verified on an AMD Strix Halo Windows box: a clean
scan now yields 22 grouped models across D:\model, D:\models,
D:\models-gguf and D:\lmstudio (drive auto-probe found the latter two on
its own), with no shards-as-models, no mmproj, and no recycle/tooling junk.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com