Math.f16round gives wrong results for some values

[ParthMozarkar]

I was going through the test262 conformance results and noticed that
Math.f16round is failing the value-conversion.js test.

After digging into it, I found that Math.f16round(1.00048828125000022204)
returns 1 instead of the correct value 1.0009765625.

The problem seems to be in how the float16 crate converts from f64 to f16
internally — it goes through f32 first, which causes double rounding and
produces the wrong result for certain inputs.

The fix would be to use the half crate's f16::from_f64 which is already
in the dependency tree and does the conversion correctly without going
through f32.

[ParthMozarkar]

could you please assign this isssue to me? i have already fixed the issue and raised a draft pr upon it