feat: dequantize MLX 4/8-bit checkpoints in the loader (#449 M3 Stage 2d)

[inureyes]

Summary

The OpenXLA loader read only bf16/f16/f32, so the many MLX-quantized checkpoints (U32-packed weights + companion *.scales / *.biases) couldn't run on the XLA path, even for the already-supported Llama / Qwen2 arches. This adds affine dequantization at load (4-bit and 8-bit). Part of #449 M3 Stage 2d. (Supersedes #489, which auto-closed when its stacked base branch merged; rebased onto main.)

What changed (mlxcel-xla)

• Config gains quantization: Option<QuantConfig{bits, group_size}> from config.json. The emitted graph is unchanged (it runs in f32); only the loader differs.
• weights::dequantize_affine unpacks the U32 weight (low-order-first, 32/bits values per word) and applies q*scale + bias per group of group_size input columns (the MLX affine layout).
• load_weights dequantizes any U32 weight using its same-shard *.scales/*.biases (f16) → the [out, in] f32 the graph expects; layernorms and q/k/v biases stay f16.

Validation

• Formula vs real data (no mlx needed): dequantizing Qwen2.5-0.5B-4bit and correlating with the original bf16 weights gives 0.995+ across embed / attention / MLP (a wrong unpacking would give ~0).
• Unit tests: hand-built u32 rows dequantize to their exact expected f32 for both 4-bit and 8-bit.
• E2E, CUDA on GB10: XLA running Qwen2.5-0.5B-4bit (the loader's dequant) is token-exact (40/40) with an HF fp32 oracle on the same weights dequantized offline. Llama-3.2-1B-4bit (different arch, tied) generates coherent text → the dequant is architecture-independent.

Notes

• Dequantizing to f32 trades the on-disk storage saving for running the quantized checkpoint on the f32 graph. Fused quantized matmul (weights packed on-device) is a later optimization (Stage 3).

Refs #449.