Fix Transformer ignoring a custom encoder or decoder with no parameters

[ayaangazali]

Proposed changes

Transformer picks its submodules with custom_encoder or TransformerEncoder(...). Since Module subclasses dict, a module that holds no arrays is an empty dict and therefore falsy, so a parameterless custom encoder or decoder gets silently thrown away and the default one is built instead:

import mlx.core as mx
import mlx.nn as nn

bool(nn.Identity())   # False

t = nn.Transformer(dims=8, num_heads=2, num_encoder_layers=3, num_decoder_layers=1,
                   custom_encoder=nn.Identity())
type(t.encoder)       # TransformerEncoder, not Identity

Same for custom_decoder. It only shows up when the custom module has no parameters of its own, so a custom encoder that happens to hold a weight works fine and this looks inconsistent rather than broken. Switched both to an explicit is not None check.

Added a small test for it. It fails on main with the current truthiness check and passes with this change.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

(test_nn.py passes; happy to drop the test if you would rather keep this to the one line change)

---

context on me: freshman contributor, and i lean on Claude Code when hunting for things like this, but i confirmed the empty-module-is-falsy behaviour myself and checked that a custom module holding a parameter behaves differently, which is what made it click. if you would rather express the fix as a conditional expression than an if/else, say the word.