perf: Optimize regular numbers parse logics

[filzrev]

This PR intended to optimize performance for number deserialization by changing Parse to TryParse method.
It's discussed at #794

Additionally this PR fix double value deserialization problems that are found on UnitTests.

What's Changes

• Replace existing regular number Parse logics with TryParse method.
• Separate float/double parse logics for .NET Framework.
• Add code path for .NET that parse text as double, then checks float value ranges.
• Add related unit tests (UnquotedStringTypeDeserialization_RegularNumbers)

.NET Behavior differences
On .NET Framework. float.Parse((double.MaxValue + 1).ToString()) throw OverflowException.
But on .NET Core. It returns PosititiveInfinite(∞).

These behavior differences are described at Single.Parse Method page.

In .NET Core 3.0 and later, values that are too large to represent are rounded to [PositiveInfinity] or
NegativeInfinity as required by the IEEE 754 specification.
In prior versions, including .NET Framework, parsing a value that was too large to represent resulted in failure.

On current implementation.
Values outside the range of the float type are converted to -∞/∞.
So I've added logics to check float value ranges.

[filzrev]

Microbenchmark results

Optimized version run about 10x faster. and memory allocation is reduced about 50%.

Baseline

Method	Job	Runtime	Mean	Error	StdDev	Exceptions	Gen0	Gen1	Allocated
Deserialize	ShortRun-.NET 8.0	.NET 8.0	176.9 us	18.02 us	0.99 us	18.0000	7.3242	-	29.96 KB
Deserialize	ShortRun-.NET Framework 4.7	.NET Framework 4.7	588.5 us	60.13 us	3.30 us	23.0000	9.7656	0.9766	43.63 KB

Optimized

Method	Job	Runtime	Mean	Error	StdDev	Exceptions	Gen0	Gen1	Allocated
Deserialize	ShortRun-.NET 8.0	.NET 8.0	20.88 us	3.234 us	0.177 us	-	4.0283	-	16.45 KB
Deserialize	ShortRun-.NET Framework 4.7	.NET Framework 4.7	52.78 us	9.504 us	0.521 us	-	5.3711	0.1221	23.66 KB

Test Code

[MemoryDiagnoser]
[ExceptionDiagnoser]
[ShortRunJob(RuntimeMoniker.Net80)]
[ShortRunJob(RuntimeMoniker.Net47)]
public class AttemptUnknownTypeDeserializationBenchmarks
{
    private static readonly IDeserializer deserializer = new DeserializerBuilder().WithAttemptingUnquotedStringTypeDeserialization().Build();

    private static readonly string TestData =
        """
        byteMaxValue:   255
        shortMaxValue:  32767
        longMaxValue :  9223372036854775807
        ulongMaxValue:  18446744073709551615
        floatMaxValue:  3.4028235E+38
        doubleMaxValue: 1.7976931348623157E+308
        """;

    [Benchmark]
    public object? Deserialize()
    {
        return deserializer.Deserialize(TestData);
    }
}

[EdwardCooke]

I like performance fixes. And this is a pretty big one! If you can address the 2 comments then I'll merge this in and get it out.

[filzrev]

Apart from this PR.
I've noticed another performance relating issue.

https://github.com/aaubry/YamlDotNet/blob/master/YamlDotNet/Serialization/NodeDeserializers/ScalarNodeDeserializer.cs#L383

Above regex checking regular number(integer or floating number).
But any text that contains digits match regex.
So unneeded Parse operations are called multiple times.

Is it able to change Regex to use more strict regex with ^ and $ ?
or is there any YAML specs that require text partial match?

[EdwardCooke]

You can add those to that regex without any issue I would think.

[filzrev]

You can add those to that regex without any issue I would think.

Thanks for your response.
I'll try to create another PR after this PR is merged.

[filzrev]

I've removed wrong comment in commit.

float.IsNormal is available .NET Core 2.1 and The behavior will be different from what I'm expecting.
https://learn.microsoft.com/en-us/dotnet/api/system.single.isnormal?view=net-8.0#remarks