Re-implement perf-optimization

[Thorium]

This is re-implementation of #68 with little bit smaller change set this time.
All tests pass, but I don't know how good is the current test-coverage.

Meanwhile it's not magic, it does cut some execution time, here are the results from this PR dbarbashov#1

BenchmarkDotNet v0.15.2, Windows 11 (10.0.26100.5074/24H2/2024Update/HudsonValley)
13th Gen Intel Core i9-13900H 2.60GHz, 1 CPU, 20 logical and 14 physical cores
.NET SDK 9.0.304
[Host] : .NET 9.0.8 (9.0.825.36511), X64 RyuJIT AVX2 DEBUG
DefaultJob : .NET 9.0.8 (9.0.825.36511), X64 RyuJIT AVX2

1. BoxedArrayRoundtrip

Method	Mean	Error	StdDev	Gen0	Gen1	Allocated	Version
Newtonsoft.Json	2.127 us	0.0412 us	0.0491 us	0.8850	0.0114	10.86 KB	13.0.1
FSharp.Json	6.076 us	0.0505 us	0.0448 us	0.9460	0.0076	11.63 KB	master
FSharp.Json	4.340 us	0.0398 us	0.0372 us	0.9384	0.0076	11.53 KB	this-PR

5. FSharpBinaryRoundtrip

Method	Mean	Error	StdDev	Gen0	Gen1	Gen2	Allocated	Version
Newtonsoft.Json	13.04 ms	0.211 ms	0.197 ms	1484.3750	984.3750	125.0000	17.92 MB	13.0.1
FSharp.Json	49.42 ms	0.461 ms	0.408 ms	2700.0000	500.0000	-	33.27 MB	master
FSharp.Json	47.87 ms	0.203 ms	0.180 ms	2800.0000	400.0000	-	33.73 MB	this-PR

6. FSharpListRoundtrip

Method	Mean	Error	StdDev	Gen0	Gen1	Allocated	Version
Newtonsoft.Json	52.50 us	1.002 us	1.029 us	7.9346	0.4883	97.55 KB	13.0.1
FSharp.Json	376.12 us	3.696 us	3.457 us	64.9414	10.7422	798.8 KB	master
FSharp.Json	322.84 us	2.010 us	1.678 us	63.4766	12.2070	783.17 KB	this-PR

11. LargeTupleRoundtrip

Method	Mean	Error	StdDev	Gen0	Gen1	Allocated	Version
Newtonsoft.Json	4.137 us	0.0702 us	0.0657 us	1.0834	0.0229	13.29 KB	13.0.1
FSharp.Json	11.809 us	0.2158 us	0.1913 us	1.5869	-	19.82 KB	master
FSharp.Json	9.760 us	0.1667 us	0.1559 us	1.5869	-	19.63 KB	this-PR