[llvm] [InstCombine] Restore narrowing of double to float for integer casts (PR #190550)

[Kavin Gnanapandithan via llvm-commits]

KavinTheG wrote:

@SavchenkoValeriy, would appreciate a review. 

I'm not sure if I interpreted your comment correctly in making `bool canBeCastedExactlyIntToFP(Value &V, Type *FP)`
and having `isKnownExactCastIntToFP` get a source of cast and the target of cast and pass it into it. 

Without the `CastInst &I`, `canBeCastedExactlyIntToFP` can't perform the KnownBit analysis, so I left it in `isKnownExactCastIntToFP`.

https://github.com/llvm/llvm-project/pull/190550