Optimize PreComputeTupleConstructor

[kerams]

BenchmarkDotNet=v0.12.1, OS=Windows 10.0.19042
AMD Ryzen 7 3700X, 1 CPU, 16 logical and 8 physical cores
.NET Core SDK=5.0.200-preview.21077.7
  [Host]     : .NET Core 5.0.2 (CoreCLR 5.0.220.61120, CoreFX 5.0.220.61120), X64 RyuJIT DEBUG
  DefaultJob : .NET Core 5.0.2 (CoreCLR 5.0.220.61120, CoreFX 5.0.220.61120), X64 RyuJIT

Method	Mean	Error	StdDev	Median	Gen 0	Gen 1	Gen 2	Allocated
New3Part	134.50 ns	3.313 ns	9.768 ns	129.21 ns	0.0842	-	-	704 B
Old3Part	1,758.37 ns	19.991 ns	18.699 ns	1,757.57 ns	0.1755	-	-	1480 B
New1Part	14.62 ns	0.330 ns	0.309 ns	14.68 ns	0.0115	-	-	96 B
Old1Part	283.67 ns	2.794 ns	2.614 ns	282.73 ns	0.0162	-	-	136 B

Tested against

let threePart = (("yup", 1s), 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, Some 12, 13, "nope",  17, 18, ValueSome 19)
let onePart = ("yup", 1s)

There's one little problem though. When values of the wrong type are passed to the precomputed constructor, InvalidCastException is now thrown instead of ArgumentException (MakeTuple still throws this as well). I'm not sure if this matters that much (I suspect a similar change sneaked into #9784), but if it does, I could theoretically emit a runtime check throwing the old exception here at the cost of a small performance hit.

[cartermp]

This is technically a breaking change given the change in exception that's thrown, but I think that's a very minor thing given the perf benefits.

[cartermp]

Could you update the failing test?