set default value of mlp_impl to grouped in layers/arguments

[comane]

What does this PR do?

Sets the default value of mlp_impl to grouped in megablocks/layers/arguments/Arguments dataclass.

This is done to avoid the default dmoe script in exp/dmoe/dmoe_46m_8gpu.sh to crash with:

[rank0]:   File "/home/mark/codes/megablocks/megablocks/layers/arguments.py", line 82, in __post_init__
[rank0]:     raise ValueError(
[rank0]: ValueError: Sparse MLP is not supported with triton >=3.2.0. Please use mlp_impl="grouped" instead.

What issue(s) does this change relate to?

Fixes #170

Before submitting

• Have you read the contributor guidelines?
• Is this change a documentation change or typo fix? If so, skip the rest of this checklist.
• Was this change discussed/approved in a GitHub issue first? It is much more likely to be merged if so.
• Did you update any related docs and document your change?
• Did you update any related tests and add any new tests related to your change? (see testing)
• Did you run the tests locally to make sure they pass?
• Did you run pre-commit on your change? (see the pre-commit section of prerequisites)