Update `mlx-swift-lm` dependency to 3.x to support newer model architectures (Gemma 4, etc.)

[james-333i]

Summary

AnyLanguageModel 0.8.0 pins mlx-swift-lm to from: "2.25.5", which resolves to the 2.x line (<3.0). New model architectures are now landing only in mlx-swift-lm 3.x, which is unreachable from the current dependency range, so AnyLanguageModel consumers can't run them.

Evidence (mlx-swift-lm releases)

• 2.30.6 (18 Feb) — added Qwen3-VL (dense)
• 2.31.3 (1 Apr) — added Qwen3.5 (+ MoE), text & vision
• 3.31.3 (15 Apr) — first 3.x release (breaking API); adds Gemma 4 (text, vision, MoE / E2B & E4B edge variants)

So Qwen3.5 and Qwen3-VL are reachable today (we pin mlx-swift-lm 2.31.3, within AnyLanguageModel 0.8.0's range and it works great). But Gemma 4 — and any future architecture added only on the 3.x line — is blocked, because 3.x has breaking changes that 0.8.0 wasn't built against.

Request

Update AnyLanguageModel's mlx-swift-lm dependency to 3.x (adapting to the 3.x API changes) and cut a release, so newer model families (Gemma 4 first) become available to consumers. More broadly: keeping the mlx-swift-lm dependency current would let new architectures work without each one waiting on an AnyLanguageModel bump.

Happy to test a branch/pre-release against a real iOS app (on-device, multi-turn) if that helps.