Faster floating-point finiteness checks

[HertzDevil]

When checking whether a Float32 or Float64 is finite, both in Crystal code itself via #finite? and in certain primitive conversions like Float64::MAX.to_f32, the check is done with x != x * 2. It seems this could generate suboptimal code compared to direct comparisons against infinity: https://godbolt.org/z/zzT8va1oP

There is exactly one binary representation of each signed infinity for Float32 and Float64, and it should be possible to use them directly. Additionally, the llvm.is.fpclass.* intrinsic was introduced in LLVM 15, presumably to handle cases where a floating-point type has more binary representations. We should switch to this intrinsic in both the standard library and the primitive codegen if the LLVM version is new enough.