Trim ~255 MB of unused packages from Studio setup (#4395)

* Comment out large unused packages from Studio setup requirements

Audited all packages installed by `unsloth studio setup` against actual
imports in unsloth, unsloth_zoo, and studio/backend. The following have
zero imports anywhere and are the largest offenders by disk size:

- gradio (148 MB) in studio.txt -- Studio uses React + FastAPI, not Gradio
- executorch (41.5 MB) in extras-no-deps.txt -- no imports found
- scikit-learn (31.8 MB) in extras.txt -- no imports found
- MeCab (19.9 MB) in extras.txt -- Japanese tokenizer, no imports found
- coremltools (10.2 MB) in extras.txt -- Apple CoreML, no imports found
- uroman (4.0 MB) in extras.txt -- romanization tool, no imports found

Total savings: ~255 MB (~32% of the 805 MB installed by setup).

Each line is commented out with the package size annotated so they can be
re-enabled easily if needed in the future.

* Restore scikit-learn -- needed by sentence_transformers

sentence_transformers is installed with --no-deps in extras-no-deps.txt,
so its sklearn dependency is not auto-resolved. Multiple modules in
sentence_transformers import sklearn at the top level (evaluation,
util/similarity), so removing scikit-learn would break embedding jobs.

studio/backend/requirements/extras-no-deps.txt

@@ -8,7 +8,7 @@ snac
 # TRL and related packages
 trl==0.23.1
 git+https://github.com/meta-pytorch/OpenEnv.git
-executorch>=1.0.1
+# executorch>=1.0.1               # 41.5 MB - no imports in unsloth/zoo/studio
 torch-c-dlpack-ext
 sentence_transformers==5.2.0
 transformers==4.57.1

studio/backend/requirements/extras.txt

@@ -4,7 +4,7 @@ tomli-w
 
 # ExecuTorch dependencies
 ruamel.yaml
-coremltools
+# coremltools                    # 10.2 MB - Apple CoreML, no imports in unsloth/zoo/studio
 expecttest
 flatbuffers
 hydra-core
@@ -15,7 +15,7 @@ pytest<9.0
 pytest-json-report
 pytest-rerunfailures==15.1
 pytest-xdist
-# Also needed by sentence_transformers
+# Also needed by sentence_transformers (installed with --no-deps in extras-no-deps.txt)
 scikit-learn==1.7.1
 
 # Additional extras
@@ -26,8 +26,8 @@ omegaconf
 einx
 pyloudnorm
 openai-whisper
-uroman
-MeCab
+# uroman                         # 4.0 MB - romanization, no imports found
+# MeCab                          # 19.9 MB - Japanese tokenizer, no imports found
 loguru
 flatten_dict
 ffmpy

studio/backend/requirements/studio.txt

@@ -10,7 +10,7 @@ datasets==4.3.0
 pyjwt
 easydict
 addict
-gradio>=4.0.0
+# gradio>=4.0.0                  # 148 MB - Studio uses React + FastAPI, not Gradio
 huggingface-hub==0.36.2
 structlog>=24.1.0
 diceware