Optimize Half.Equals

[huoyaoyuan]

Fixes #41329

Benchmark:
(See full benchmark at #41329 (comment))

Method	Job	left	right	Mean	Error	StdDev	Ratio	RatioSD	Code Size
Equal	PR	0	-0	0.4526 ns	0.0127 ns	0.0112 ns	1.00	0.00	75 B
Equal	master	0	-0	1.8901 ns	0.0135 ns	0.0126 ns	4.18	0.12	156 B
Equal	PR	65500	-65500	0.6604 ns	0.0088 ns	0.0082 ns	1.00	0.00	75 B
Equal	master	65500	-65500	1.6911 ns	0.0131 ns	0.0116 ns	2.56	0.03	156 B
Equal	PR	65500	0	0.6683 ns	0.0042 ns	0.0038 ns	1.00	0.00	75 B
Equal	master	65500	0	1.6832 ns	0.0076 ns	0.0068 ns	2.52	0.02	156 B
Equal	PR	65500	65500	0.2071 ns	0.0095 ns	0.0089 ns	1.00	0.00	75 B
Equal	master	65500	65500	2.2191 ns	0.0173 ns	0.0162 ns	10.73	0.43	156 B
Equal	PR	65500	NaN	0.6647 ns	0.0081 ns	0.0076 ns	1.00	0.00	75 B
Equal	master	65500	NaN	1.4638 ns	0.0015 ns	0.0014 ns	2.20	0.02	156 B
Equal	PR	NaN	0	0.4603 ns	0.0009 ns	0.0008 ns	1.00	0.00	75 B
Equal	master	NaN	0	2.4231 ns	0.0039 ns	0.0036 ns	5.26	0.01	156 B
Equal	PR	NaN	NaN	0.2098 ns	0.0017 ns	0.0016 ns	1.00	0.00	75 B
Equal	PR	NaN	NaN	0.4320 ns	0.0041 ns	0.0038 ns	2.06	0.02	75 B
Equal	master	NaN	NaN	2.4281 ns	0.0078 ns	0.0069 ns	11.58	0.09	156 B
Equal	master	NaN	NaN	2.0738 ns	0.0086 ns	0.0076 ns	9.89	0.09	156 B

Further question

Should we apply more AggressiveInlineing on Half members?
If we've get hardware acceleration, the implementation should change massively, and can be considered later.
Should call size bloat be considered?

[Dotnet-GitSync-Bot]

I couldn't figure out the best area label to add to this PR. If you have write-permissions please help me learn by adding exactly one area label.

[jkotas]

AggressiveInlining does improve microbenchmark results, but it introduces code bloat that tends to hurt real workloads. We are careful about using it.

Given that nothing else is annotated with AggressiveInlining in Half, I would not introduce one-off here. I think we would need some real world examples to justify it, and likely apply it to number of other methods in this file.

[huoyaoyuan]

Sure. Removed for this PR. Let it be a part of optimizing the whole Half type.

[jkotas]

Thanks