fix: set processor_repo to vision weights repo in Kimi model cards

[Drifter4242]

Just some small changes to the vision model cards. Two things: if the model isn't loaded then specifying it in the processor_repo can cause it to fail (I must have had the models downloaded before so it wasn't a problem). Also I've renamed my hf depot (old links should still work though), so I updated the links.

The rest is written by Sonnet (reviewed by me):

The processor_repo fields for Kimi K2.5 and K2.6 pointed to the full LLM repos (moonshotai/Kimi-K2.5, moonshotai/Kimi-K2.6), which are not downloaded locally (595+ GB). The image processor must be loaded from the vision weights repo, which IS downloaded alongside the vision weights.

• mlx-community/Kimi-K2.5: processor_repo -> Drifter4242/Kimi-K2.5-vision
  (also fixes weights_repo: davehind/ -> Drifter4242/ after account transfer)
• mlx-community/Kimi-K2.6-mlx-DQ3_K_M-q8: processor_repo -> exolabs/Kimi-K2.6-vision
• moonshotai/Kimi-K2.6: processor_repo -> exolabs/Kimi-K2.6-vision

Without this fix vision loading fails immediately with a FileNotFoundError when trying to load the tokenizer/processor from the non-existent local path for the full model.

Motivation

The processor_repo fields for three Kimi vision model cards pointed to the
full LLM repos (moonshotai/Kimi-K2.5, moonshotai/Kimi-K2.6). These are
500–660 GB models that are not downloaded locally. When exo loads the vision
pipeline, it calls build_model_path(ModelId(processor_repo)) to construct a
local filesystem path — if that path doesn't exist, loading fails.

The image processor must be loaded from the vision weights repo, which is
downloaded automatically alongside the vision weights and is always present.

A secondary fix: mlx-community/Kimi-K2.5 had weights_repo = "davehind/Kimi-K2.5-vision",
a now-stale HuggingFace account that was transferred to Drifter4242.

Changes

Model card	Field	Before	After
mlx-community/Kimi-K2.5	weights_repo	davehind/Kimi-K2.5-vision	Drifter4242/Kimi-K2.5-vision
mlx-community/Kimi-K2.5	processor_repo	moonshotai/Kimi-K2.5	Drifter4242/Kimi-K2.5-vision
mlx-community/Kimi-K2.6-mlx-DQ3_K_M-q8	processor_repo	moonshotai/Kimi-K2.6	exolabs/Kimi-K2.6-vision
moonshotai/Kimi-K2.6	processor_repo	moonshotai/Kimi-K2.6	exolabs/Kimi-K2.6-vision

Why It Works

vision.py resolves processor_repo via build_model_path(ModelId(processor_repo)),
which maps a HuggingFace repo ID to a local cache path. The full LLM repos are
never fully downloaded by vision users, so their local paths do not exist.
The vision-specific repos (Drifter4242/Kimi-K2.5-vision, exolabs/Kimi-K2.6-vision)
are small (~1–2 GB) and are downloaded as part of the normal vision weight fetch,
so their local paths are always present by the time the processor is needed.

Test Plan

Manual Testing

Hardware: 2× Mac Studio M3 Ultra 512 GB, Thunderbolt 5 direct bridge,
MlxJaccl RDMA tensor-parallel (moonshotai/Kimi-K2.6, 595 GB INT4, 61 layers).

• Confirmed vision pipeline loads successfully after this change with
  moonshotai/Kimi-K2.6 + exolabs/Kimi-K2.6-vision weights.
• Sent image requests and received correct responses.

Automated Testing

No code changes — model card data files only. All existing tests pass:
pytest src -m "not slow" --import-mode=importlib