* fix smoothL1Loss

* Update Source/MLXNN/Losses.swift

Co-authored-by: Awni Hannun <awni.hannun@gmail.com>

* memory limit fix
* make DeviceInfo publicly accessible

- per change in #93

* consistent treatment of HasDType (Float.self) vs dtype (.float32)

- fixes #199
- there were inconsistent methods for creating MLXArrays with various patterns using Float.self vs .float32
- add an FInfo struct also in support of dtype
- encountered when porting VLMs

Co-authored-by: Awni Hannun <awni@apple.com>

* various Random improvements

- allow a Key or a RandomState to be used with Random functions
- users can create RandomState
- users can do withRandomState to scope the use of a RandomState -- useful with multiple threads

- fixes #225
- although CompiledFunction was marked as unchecked(Sendable) it really wasn't
- guard the state with a lock

- adopt mlx v0.25.1 and mlx-c v0.2.0

- add saveToData and load from data (safetensors) -- fix #214
- fixes #211 (moves metalKernel API back to what it was)

- add withError and withErrorHandler error handling
- add function import/export: https://ml-explore.github.io/mlx/build/html/usage/export.html
 
- use an evalLock for eval, asyncEval and Stream creation

…der compiler (#231)

- switch to data.withUnsafeBytes

…229)

* Add quantize function that can return different parameters per layer
Co-authored-by: Awni Hannun <awni@apple.com>