[CoreML] Update Conv and Softmax ops (#24594)
### Description
Moved the dimension limit because it seems to only apply to conv
operations (texture memory is typically used for conv operations in the
GPU because it has a slow write but fast read -- ChromaDB model had a
slice operation with an input > 16384 -- operation worked fine after I
had moved the dim check)
Also added extra checks for Softmax on MLProgram that allows more
softmax nodes to be moved to CoreML
