Fix broken code examples in custom model docs

[singularity-14]

Both runnable examples in docs/source/evaluating-a-custom-model.mdx fail as written.

CLI example

The task argument is missing its closing quote, so the shell swallows the following --max-samples 10 line into the string literal:

 lighteval custom \
     "google-translate" \
     "examples/custom_models/google_translate_model.py" \
-    "wmt20:fr-de \
+    "wmt20:fr-de" \
     --max-samples 10

Python API example

tasks=truthfulqa:mc is a bare identifier followed by a colon, which is a SyntaxError — the example cannot run at all:

 pipeline = Pipeline(
-    tasks=truthfulqa:mc,
+    tasks="truthfulqa:mc",
     pipeline_parameters=pipeline_params,
     evaluation_tracker=evaluation_tracker,
     model_config=model_config
 )

Notes

Both task names are valid — wmt20:fr-de in src/lighteval/tasks/tasks/sacrebleu.py and truthfulqa:mc in src/lighteval/tasks/tasks/truthfulqa.py — so only the quoting was wrong, and the examples' intent is unchanged.

Verified by parsing every python block in the file with ast.parse and tokenizing every bash block with shlex.split; both now pass. The CLI example tokenizes to the three positional arguments custom expects (model_name, model_definition_file_path, tasks) plus --max-samples 10.

Relation to #760

Partially addresses #760. The imports originally reported there look already fixed on main (ParallelismManager is imported, and EnvConfig is gone); what remained was the examples still not running, which is what this PR fixes.

The base-class concern raised later in that thread is a separate, larger change and isn't touched here — so I'd suggest leaving #760 open to track it, unless you'd rather close it.