CpuMath Enhancement: Double-compute input elements in hardware intrinsics

Style changes needed to solve part of https://github.com/dotnet/machinelearning/issues/823

After implementing "double-compute", it is expected to make hardware intrinsics more efficient.

## Details (mostly from @tannergooding)
- In `src\Microsoft.ML.CpuMath\SseIntrinsics.cs` and `src\Microsoft.ML.CpuMath\AvxIntrinsics.cs`, change the last loop of the existing 3-loop code pattern into the following:
1. Saving the stored result (`dstVector`) from the last iteration of the vectorized code
2. Moving `pDstCurrent` back such that `pDstCurrent + elementsPerIteration == pEnd`
3. Doing a single iteration for the remaining elements
4. Mix the saved result from the last iteration of the vectorized code with the result from the remaining elements
5. Write the result

This generally results in more performant code, depending on the exact algorithm and number of remaining elements

- On handling unpadded parts in AVX intrinsics:

For some algorithms (like `Sum`), it is possible to “double-compute” a few elements in the beginning and end to have better overall performance. See the following pseudo-code:
```
if addr not aligned
tmp = unaligned load from addr
tmp &= mask which zero's elements after the first aligned address
result = tmp
move addr forward to the first aligned address

while addr is aligned and remaining bits >= 128
result += aligned load
addr += 128-bits

if any remaining
addr = endAddr - 128
tmp = unaligned load from addr
tmp &= mask which zero's elements already processed
result += tmp

Sum the elements in result (using "horizontal add" or "shuffle and add")
```

So, your overall algorithm will probably look like:
```
if (Avx.IsSupported && (Length >= AvxLimit))
{
// Process 256-bits, we have a limit since 256-bit
// AVX instructions can cause a downclock in the CPU
// Algorithm would be similar to the SSE pseudo-code
}
else if (Sse.IsSupported && (Length >= SseLimit))
{
// Pseudo-code algorithm given above

// 128-bit instructions operate at full frequency
// and don't downclock the CPU, we can only use
// them for more than 128-bits so we don't AV
}
else
{
// Software Implementation
}
```

If you can’t “double-compute” for some reason, then you generally do the “software” processing for the beginning (to become aligned) and end (to catch stray elements).
• `AvxLimit` is generally a number that takes into account the “downclocking” that can occur for heavy 256-bit instruction usage
• `SseLimit` is generally 128-bits for algorithms where you can “double-compute” and some profiled number for other algorithms

cc: @tannergooding since he suggested this approach.