Fix checkpointing docs: Loading a comma-seprated string from toml file is off (#1281)

As title

From the test:
https://github.com/pytorch/torchtitan/blob/main/tests/unit_tests/test_job_config.py#L123
- Commandline supports comma-seprated string,
- TOML file only support TOML native syntax for `list[str]`:
`exclude_from_loading = ["optimizer", "lr_scheduler", "dataloader"]`

Author: wwwjn

docs/checkpoint.md

@@ -64,12 +64,13 @@ python -m torch.distributed.checkpoint.format_utils dcp_to_torch torchtitan/outp
 
 7. EXCLUDING SPECIFIC KEYS FROM CHECKPOINT LOADING
 In some cases, you may want to partially load from a previous-trained checkpoint and modify certain settings, such as the number of GPUs or the current step. To achieve this, you can use the `exclude_from_loading` parameter to specify which keys should be excluded from loading.
-This parameter takes a comma-separated list of keys that should be excluded from loading.
+This parameter takes a list of string that should be excluded from loading.
 ```
 [checkpoint]
 enable_checkpoint = true
-exclude_from_loading = "data_loader,lr_scheduler"
+exclude_from_loading = ["data_loader", "lr_scheduler"]
 ```
+When used in command line, the parameter should be a comma-separated list of strings. For example: `--checkpoint.exclude_from_loading data_loader,lr_scheduler`.
 
 That's it. You have now successfully converted a sharded torchtitan checkpoint for use in torchtune.