cc: fix aarch64 float16 conversion using fmov instead of fcvt

emit_float_to_float() only recognized Float↔Double conversions in its
needs_convert check, causing Float16↔Float/Double conversions to emit
fmov (bit-copy) instead of fcvt (type conversion). The float16 bit
pattern was zero-extended to 32 bits rather than properly converted,
producing wrong values for all float16 arithmetic on aarch64.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: jgarzik

cc/arch/aarch64/float.rs

@@ -587,6 +587,14 @@ impl Aarch64CodeGen {
         let src_kind = insn.src_typ.map(|t| types.kind(t));
         let dst_kind = insn.typ.map(|t| types.kind(t));
         let needs_convert = match (src_kind, dst_kind) {
+            (
+                Some(TypeKind::Float16),
+                Some(TypeKind::Float | TypeKind::Double | TypeKind::LongDouble),
+            ) => true,
+            (
+                Some(TypeKind::Float | TypeKind::Double | TypeKind::LongDouble),
+                Some(TypeKind::Float16),
+            ) => true,
             (Some(TypeKind::Float), Some(TypeKind::Double | TypeKind::LongDouble)) => true,
             (Some(TypeKind::Double | TypeKind::LongDouble), Some(TypeKind::Float)) => true,
             // On aarch64, Double and LongDouble are the same, no conversion needed