fix(onnx): use explicit max_width in SpanMLP.view() for ONNX export

[ajroetker]

Summary

Fixes ONNX export failure for models using SpanMLP span representation.

The SpanMLP class used an inferred dimension (-1) in the view() call:

span_rep = span_rep.view(B, L, -1, D)

This causes ONNX export to fail with a reshape size mismatch error because PyTorch's ONNX exporter cannot reliably infer the -1 dimension when using dynamic shapes.

Error example:

total requested size 786432 (dimensions=[1 128 12 512]) doesn't match original size 65536 (dimensions [1 128 512])

Fix

1. Store max_width as an instance variable in __init__
2. Use self.max_width explicitly in view() instead of -1

# Before
span_rep = span_rep.view(B, L, -1, D)

# After  
span_rep = span_rep.view(B, L, self.max_width, D)

This makes SpanMLP consistent with other span representation classes (SpanMarker, SpanMarkerV0, SpanMarkerV1) which already use explicit max_width in their view() calls.

Testing

• This fix was discovered while attempting to run exported GLiNER ONNX models in GoMLX
• The fix allows successful ONNX export and inference with dynamic batch sizes and sequence lengths