Some System.Decimal performance improvements (#99212)

* Improve performance of decimal division

* Remove usage of mulx via instrinct on 32bit x86 since it produces worse code than

Job=ShortRun  IterationCount=3  LaunchCount=1
WarmupCount=3

| Method                     | a | b          | Mean     | Error     | StdDev    | Allocated |
|--------------------------- |-- |----------- |---------:|----------:|----------:|----------:|
| Mul64By32_New | 3 | 4294967295 | 2.068 ns | 0.0459 ns | 0.0383 ns |         - |
| Mul64By32_Ori                |  3   | 4294967295 | 2.916 ns | 0.0231 ns | 0.0193 ns |         - |

* update part of Multiply

- Add comment to BigMul64By32 and make it return nunit to avoid clearing upper 32 bits
- Simplify IncreaseScale

* remove unintentional file

* Improve division by 64bit value on x64

* Remove some more 64bit divides for x86

* Call IncreaseScale in one more place

* add extra paranthesis

* review: remove X86.X86Base.X64.DivRem

* Remove BigMul64By32 and add overloads of Math.BigMul instead so that it can easily be removed once JIT recognize and optimize "ulong * uint"

* Simplify Div96By32

* Remove 64 bit path from Div96By32

* Add Div64By32 helper to avoid check for X86 in multiple places

* Use 64*32 multiply in more places

* Add new IncreaseScale overload to fix issue with Store-To-Load forwarding
* Gives around 20% faster perf for 64bit for full 96bit division
* removes cast

* Add back old 32bit code for IncreaseScale
* makes the 5-10% regression 20% faster (25% speedup) for 1/3

* 6% faster with longer increasescale

* Add back  Bmi2.MultiplyNoFlags in Math.BigMul

* fix Math.Bigmul compilation

* switch from Bigmul to (ulong)a * (uint)b in IncreaseScale
* Gives up to 10ns (or 14% faster)

* Call IncreaseScale(ref Buf12) from IncreaseScale(ref Buf16) for 32bit code

* Add #pragma warning disable CA2252

* Change supression to SYSLIB5004 now that DivRem is marked as [Experimental]

---------

Co-authored-by: Tanner Gooding <[email protected]>

Author: Daniel-Svensson