[arm64] Tune memory copy performance

[jkotas]

Check what was done for x86/x64 in dotnet/coreclr#9786

[sdmaclea]

@jkotas @jashook Please assign this to me.

If I look at the benchmark which was used to optimize dotnet/coreclr#9786, it looks flawed. All the measurements are made with a fairly small bucket. This will train the branch predictors and result in excessively branchy code. I think that is why the "optimized" Buffer.cs code is so complicated. I will add some additional tests which will hopefully detect the branch prediction penalties.

[jkotas]

cc @vkvenkat