[llvm] [ValueTracking] Improve `Bitcast` handling to match SDAG (PR #125935)

[Simon Pilgrim via llvm-commits]

================
@@ -1304,6 +1304,9 @@ static void computeKnownBitsFromOperator(const Operator *I,
     // Look through a cast from narrow vector elements to wider type.
     // Examples: v4i32 -> v2i64, v3i8 -> v24
     unsigned SubBitWidth = SrcVecTy->getScalarSizeInBits();
+    unsigned NumElts = DemandedElts.getBitWidth();
+    unsigned SubScale = BitWidth / SubBitWidth;
----------------
RKSimon wrote:

SubScale isn't the same for each path - the new path is `SubBitWidth / BitWidth`

https://github.com/llvm/llvm-project/pull/125935