Let conversion ops declare many-to-many support, and fix four loader details
- ConversionOps.supports_many_to_many replaces the hard-coded tuple of Ernie
4.5 VL ops that WeightConverter checked, so core imports no op to allow it.
- SplitModulelist squeezes only the split dim, keeping a size-1 dim of the
tensor's own (a single column of block scales).
- WeightTransform.copy keeps force_cpu.
- A param quantized on load is no longer cast to the empty param's dtype
first (int8 storage would zero it, float8 would double-round it).
- revert_weight_conversion applies renames after the converters, so a
converted key is resolved under the model's names before it is renamed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>