chore(ios): update MLX backend binary with compiled-execution path

[NorbertKlockiewicz]

Description

Updates the prebuilt iOS MLX backend binary (libbackend_mlx_ios.a and the ExecutorchLib.xcframework that embeds it) with the latest ExecuTorch MLX backend.

New backend capabilities baked into the binary:

• RMSNorm fusion — fuses the Gemma-style fp32-upcast RMSNorm cluster into a single fused op.
• mx::compile compiled-execution path — traces the interpreter once per input signature and replays the fused graph, cutting per-token Metal kernel encoding.
• rope_t custom op — tensor-offset RoPE so the decode graph has no .item() (required for mx::compile).
• Tensor-based index_copy — KV-cache update without an eval-during-trace sync.
• Copy-mode constant loading + residency guard — materializes weights into MLX-owned memory and fails loudly if the copy is elided.

This binary is required to run the compile-friendly Gemma 4 E2B MLX .pte exports; the previous binary would not execute them correctly.

Introduces a breaking change?

• Yes
• No

Type of change

• Bug fix (change which fixes an issue)
• New feature (change which adds functionality)
• Documentation update (improves or adds clarity to existing documentation)
• Other (chores, tests, code style improvements etc.)

Tested on

• iOS
• Android

Testing instructions

1. Build and run the LLM example app on an iOS device.
2. Load a compile-friendly Gemma 4 E2B MLX model and generate text.
3. Confirm generation is coherent and runs at the improved decode rate (the runtime logs compiled execution: enabled on first inference).

Screenshots

Related issues

Checklist

• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have updated the documentation accordingly
• My changes generate no new warnings

Additional notes

This PR contains only the rebuilt binary artifacts. The corresponding backend source lives in the ExecuTorch fork; this updates the prebuilt that the iOS app links against. The simulator variant (libbackend_mlx_simulator.a) is unchanged in this PR.

This PR was authored with Claude.