fix nn.GRU skipping bhn bias when hidden is None - #3252

Proposed changes

#3249

when hidden=None (the default), the GRU new-gate computation skips the
hidden-side bias bhn entirely. the first timestep computes:

n = tanh(W_xn·x + b_n)

but with an implicit zero hidden state it should compute:

n = tanh(W_xn·x + b_n + r ⊙ b_hn)

this makes gru(x) produce different results from
gru(x, hidden=mx.zeros(...)) whenever bias=True.

the fix adds an elif branch that applies r * self.bhn when there is
no hidden state but the bias exists.

Checklist

• 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)