Add invSqrt

[jmcwilliams403]

This is the original Quake III Arena implementation of Fast InvSqrt() from id Tech 3, ported to Java using Float.floatToIntBits() and Double.doubleToLongBits(). May need to add a disclaimer that any observable performance difference over 1/Math.sqrt(x) is entirely system-dependent

[tommyettinger]

I just want to point out why this was closed: the method is sound and should work for the -0.5 power with good precision, but according to benchmarks I ran a while ago (partly in https://github.com/tommyettinger/assorted-benchmarks/blob/master/src/main/java/net/adoptopenjdk/bumblebench/examples/MathSqrtFloatBench.java ), this inverse approximation isn't as fast as 1f/(float)Math.sqrt(x) because sqrt is an intrinsic candidate on most JDKs. For cube root, an approximation like this is much faster than Math.cbrt(x), so the format is still useful, just not the specific case of the 0.5 or -0.5 power.

[tommyettinger]

So my old benchmark didn't test GraalVM (it wasn't usable on Windows for a while early on), and it turns out this code is significantly faster on Graal (> 40% higher throughput) while only being a tiny bit slower elsewhere (less than 5% slower on HotSpot, though since it's an approximation, it is less accurate). It looks like Graal doesn't have the same intrinsic optimizations that HotSpot has for Math.sqrt(), or they aren't as fast. I can't reopen this PR because of some Git reason ("main branch was force-pushed or recreated"), so I may recreate it with a notice about being a good choice if you distribute with or typically target GraalVM. There was one change I needed to make anyway; Float's int/float conversions are good on Desktop and Android, but are extremely slow by default on GWT. We use BitConversion here, which should inline to be the same as Float.intBitsToFloat() and related methods on desktop, while using an alternate (non-approximate) algorithm on GWT-only that is optimized for what it can do.