feat: fused Metal q4 inference for MLX 4-bit models

[cjchanh]

What

This patch enables Cake to serve MLX 4-bit quantized models that upstream main cannot load. On M5 Max, Qwen2.5-7B-Instruct-4bit loaded at 9.5 GiB and generated 10 tokens at 56.71 tok/s in a bounded API run.

How

• q4_matvec_f16 and q4_matmul_tiled_f16 MSL kernels that read packed 4-bit weights and dequantize on-the-fly during matmul
• QuantizedLinear layer type that stores packed U32 weights + F16 scales/biases on Metal without expansion
• MetalMlxBackend VarBuilder that auto-detects MLX 4-bit format and keeps weights packed
• MLP and Attention layers use polymorphic LinearWeight (Dense/Quantized)
• Non-quantized tensors (embeddings, norms, lm_head) fall through to standard F16 dequantization

Tested

• cargo test -p cake-core --features metal
• cargo clippy -p cake-core -p cake-cli --features metal
• Verified on M5 Max 128GB: 56.71 tok/s, coherent output
• Verified on M1 Air 8GB: model loads at 1.5 GiB (memory-constrained throughput due to candle buffer pool growth, see Metal buffer pool grows beyond physical RAM on small unified-memory devices huggingface/candle#3464)

Files

• cake-core/src/backends/metal/ops.msl — MSL kernels
• cake-core/src/backends/metal/mod.rs — kernel dispatch + CustomOp
• cake-core/src/backends/mod.rs — trait method
• cake-core/src/utils/quantized_linear.rs — QuantizedWeight + LinearWeight
• cake-core/src/utils/mlx_quant.rs — MLX detection
• cake-core/src/utils/gptq.rs — MetalMlxBackend
• cake-core/src/utils/mod.rs — auto-detection wiring
• cake-core/src/models/common/mlp.rs — quantized MLP
• cake-core/src/models/common/attention.rs — quantized attention
• cake-core/tests/unit_tests/test_quantization.rs — q4 validation tests

[cjchanh]

Additional benchmark context for this PR.

Benchmark results

M5 Max (128 GB), single-device

• model: mlx-community/Qwen2.5-7B-Instruct-4bit
• throughput: 56.71 tok/s (10 tokens in a bounded API request)
• loaded memory: 9.5 GiB
• upstream main: cannot load this checkpoint (shape mismatch for model.embed_tokens.weight, expected [152064, 3584], got [152064, 448])

M5 Max + iPad Air M3, distributed (0.5B)

• model: mlx-community/Qwen2.5-0.5B-Instruct-4bit
• iPad worker discovered via zero-config discovery
• iPad assigned model.layers.0
• throughput: 55.44 tok/s

M1 Air + iPad Air M3, distributed (0.5B)

• model: mlx-community/Qwen2.5-0.5B-Instruct-4bit
• iPad worker assigned model.layers.0 via manual topology
• throughput: 29.17 tok/s on a bounded request
• master memory at generation: 854.7 MiB

Tested on Apple Silicon (M5 Max, M1 Air) and iPad Air M3.

This is the key distinction for the patch: it does not just optimize an existing path. It enables Cake to serve MLX 4-bit checkpoints that upstream main cannot load, and it does so at usable throughput on Apple Silicon.

[cjchanh]

This PR is still relevant from my side, and I am still maintaining the branch.
I realize this is a larger review than the other pending patches because it includes the fused Metal q4 path and validation coverage.
If review size is the blocker, I can split it into smaller PRs around format detection, packed-weight storage, Metal kernels, and model integration.
I can also respond to review feedback on the current branch if that is easier.