Unable to build candle with flash attention on iOS

When I try to build and run a llama 3.2 1b model on iOS (iPhone 14) with flash attention on Metal, I get ``/Users/jpchen/.cargo/git/checkouts/candle-6740f55d69a3bf41/b4ec636/candle-transformers/src/models/llama.rs:254:5:
not implemented: compile with '--features flash-attn'`

A little unfamiliar with Candle - I see that flash attention is supported for Metal hardware, and I was curious if this is an ios specific thing or if theres a way I could build it to get flash attention support? Thanks.

Start at candle-transformers/src/models/llama.rs around line 254 and trace how the flash-attn feature is enabled for Metal builds. Reproduce the iOS build with the llama 3.2 1b model, then determine whether flash attention is supported on iOS and what a supported build path requires. Done means the iOS behavior and required configuration are established.

Written by the indexing model from the issue text.