[Skia] RenderTargetBitmap improvements

[Youssef1313]

Current situation

• RenderTargetBitmap is IDisposable in Uno while it's not in WinUI. We should maintain parity with WinUI.
• RenderTargetBitmap is using ArrayPool for renting the array. Generally, RenderTargetBitmap will tend to ask for a very large array that the .NET ArrayPool implementation will just always allocate a new array. So, effectively we are always allocating a large array and not using anything from the pool.
• The allocated array ends up in the Large Object Heap (which might be a performance issue)

Two solutions come to mind:

Solution 1: Avoid _buffer field

Currently, in RenderAsBgra8_Premul, we get an SKBitmap and copy its pixels to _buffer field.

Then, we dispose the SKBitmap, and we use the _buffer array for two purposes:

1. Creating an SKBitmap back again, if needed (e.g, if RenderTargetBitmap is used as Image.Source)
2. In GetPixelsAsync

Both purposes can still be achieved with SKBitmap instead of a _buffer array.

The remaining part will be disposing the SKBitmap, which we will have to do in the destructor (unfortunately)

Note that this approach is Skia-specific. We might find similar ideas for other platforms, but then this will introduce more platform-specific code to RenderTargetBitmap.

EDIT: I tried this approach, but it's problematic as we need the array anyways in GetPixelsAsync (because Buffer needs either an array of bytes or Memory<byte>, but we have a Span<byte> that we can't convert to Memory<byte> as far as I know.

Solution 2: Using SegmentedArrays

We can port SegmentedArray implementation from Roslyn and use it instead of a regular array. This will help with LOH allocations.

EDIT: This too couldn't work, as SkiaSharp needs a pointer to contiguous memory.

---

Some profiling will be needed to know the performance impact of the changes. This will be most notable in snapshot tests.

[Youssef1313]

Both approaches didn't work. I'm experimenting now with an unmanaged memory approach (suggested by @ramezgerges)