fix float types in Phi (use float16) (#25)

- per suggestions in #23 ensure that the values that go into the cache are float16

Libraries/LLM/Phi.swift

@@ -73,14 +73,12 @@ private class PhiAttention: Module {
             keys = rope(keys)
         }
 
-        queries = queries.asType(Float.self)
-        keys = keys.asType(Float.self)
-
         // Finally perform the attention computation
         let scale = sqrt(1 / Float(queries.dim(-1)))
         let output = MLXFast.scaledDotProductAttention(
-            queries: queries, keys: keys, values: values, scale: scale, mask: mask
+            queries: queries.asType(.float32), keys: keys, values: values, scale: scale, mask: mask
         )
+        .asType(values.dtype)
         .transposed(0, 2, 1, 3)
         .reshaped(B, L, -1)