arm64: Remove widening casts before truncating

[jonathandavies-arm]

This PR remove unnecessary casts instructions. I've added some examples below taken from the Casts.cs test file but I've seen the same optimisations when I run the asmdiffs.

Example 1

[MethodImpl(MethodImplOptions.NoInlining)]
static sbyte CastIntLongSbyte(int x)
{
    //ARM64-FULL-LINE: sxtb {{w[0-9]+}}, {{w[0-9]+}}
    return (sbyte)(long)x;
}

Arm64 Instruction Diff

-sxtw    x0, w0
 sxtb    w0, w0

x64 Instruction Diff

-movsxd  rax, edi
-movsx   rax, al
 movsx   rax, dil

Example 2

[MethodImpl(MethodImplOptions.NoInlining)]
static ushort CastULongUShortULongUShort(ulong x)
{
    //ARM64-FULL-LINE: uxth {{w[0-9]+}}, {{w[0-9]+}}
    return (ushort)(ulong)(ushort)x;
}

Arm64 Instruction Diff

 uxth    w0, w0
-mov     w0, w0
-uxth    w0, w0

x64 Instruction Diff

-movzx    rdi, di
-mov      eax, edi
-movzx    rax, ax
 movzx    rax, di

[jonathandavies-arm]

@dotnet/arm64-contrib

[EgorBo]

Why do we need these if we can use xunit's Assert.Equal and friends?