Replace all buffer allocations with pools, Span<T> over R's memory

[buybackoff]

There are several places with new byte[] and new double[] that need pooling.

Also, the code like below needs modern-era rewrite with Unsafe.Read, Unsafe.CopyBlk or even SIMD (see corefxlab/Kestrel) if the IL opcode doesn't vectorize internally.

var data = new byte[sizeof(double)];
            for (int byteIndex = 0; byteIndex < data.Length; byteIndex++)
            {
                data[byteIndex] = Marshal.ReadByte(pointer, offset + byteIndex);
            }
            return BitConverter.ToDouble(data, 0);

There are probably other low-hanging fruits such as mentioned above.

Consider using Span and/or OwnedMemory backed by R's unmanaged memory.

Try to avoid changing Marshal methods to Unsafe just for the sake of it. Only obvious cases such as reading a single double via a loop over its bytes. Later could review all Marshal usages.

Review ArrayConvertAllTwoDim method in ArrayConverter. I believe there is a simpler way to convert 1D <-> 2D array using System.Array.

P.S. For Spreads, any R usage is considered slow, i.e. not an online algo, but batch processing or complex math. So better be safe and do not break things, rather profile remaining non-obvious cases and fix them ad-hoc.