fix(vlm): load non-affine VLM weights with the right quant mode and group_size

[inureyes]

Summary

Vision encoders (and any caller that does not thread an explicit quantization mode) load weights through from_weights, which hardcoded the affine mode. On block-float VLMs that aborts in quantized_matmul (Biases must be provided for affine) because mxfp4/nvfp4/mxfp8 weights carry no zero-point biases. minicpm-v-4.6-mxfp4 was the visible casualty.

Two reconciliations fix it:

1. Mode auto-detect in QuantizedEmbedding::from_weights and UnifiedLinear::from_weights: a .biases tensor means affine; its absence means a block-float scheme, distinguished by bits (8 → mxfp8) and group_size (16 → nvfp4, else mxfp4).
2. group_size reconciliation for block-float modes in UnifiedLinear::from_weights_with_mode. For affine, bits are inferred against the trusted group_size (unchanged). For block-float modes the bits are fixed by the mode, so group_size is derived from the shapes instead: in_features = packed_in * (32/bits) = num_groups * group_size. A mixed export such as minicpm-v-4.6-mxfp4 stores some weights at group_size 32 while config.json reports 64; the old loader passed 64 straight through and mismatched in the matmul.

Verification (GB10, CUDA)

Model	Before	After
minicpm-v-4.6-mxfp4 (with image)	abort (Biases must be provided for affine, then group_size mismatch)	coherent description, 139.85 tok/s decode
qwen2.5-7b-4bit (affine sanity)	53 tok/s	53 tok/s (path unchanged)
gpt-oss-20b-mxfp4 (block-float sanity)	78 tok/s	78 tok/s (inferred params equal passed)

For affine models and correctly exported block-float models the inferred params equal the passed ones, so behavior is byte-identical.