fix: resolve Gemma4 K-eq-V broadcast crash (closes #59)

[solderzzc]

Summary

Bumps the mlx-swift-lm submodule to SharpAI/mlx-swift-lm#23 which fixes the root cause of the crash reported in #59.

Root Cause

All gemma-4-26b-a4b-it-* models use attention_k_eq_v: true for full-attention layers. In this path, the Swift port had a subtle double-transpose bug:

k → kNorm → transpose → [B, nKvH, L, D]
v = k  (already transposed)
v → vNorm → transpose → [B, L, nKvH, D]  ← WRONG!

The wrongly-shaped v caused the crash at SDPA:

[broadcast_shapes] (1,512,2,512) vs (1,2,512,512)

Changes (in mlx-swift-lm)

• MLXLLM/Models/Gemma4Text.swift: When vProj is nil (K-eq-V path), skip the redundant transpose — just v = vNorm(k).
• MLXVLM/Models/Gemma4.swift (stretch): Added LayerPartitionable + StreamableMoE conformance to Gemma4TextBackbone and wired streamExperts through the public Gemma4 VLM class, enabling SSD expert streaming on the VLM path (mirrors Qwen35.swift).

Testing

• swift build --target MLXLLM ✅
• swift build --target MLXVLM ✅
• swift test --filter Gemma4Tests ✅

Closes #59

[notatestuser]

Awesome and amazing. Thank you!

[solderzzc]

Awesome and amazing. Thank you!

Thanks for your detailed bug report.