Fix precision of real32quant for 32 bit float

[undefined-panda]

No description provided.

[hahnjo]

I think this change would benefit from more details in the commit message: what does ROOT do (compute in double), what does the macro request (single-precision float so the value is cast), why is that not possible in JavaScript (no single precision float type), and how does the change solve it (Math.fround "cutting off" the mantissa).

Also you need to update jsroot/types/fundamental/README.md to remove the caveat documented so far.

[hahnjo]

regarding the commit message:

ROOT computes quantized floats in double precision but can write them with single- or double-precision.

Not quite, the quantization is eventually written as integers with a fixed number of bits. The important point is that the in-memory field type can be single- or double-precision, depending on the requested template argument.

Since JS only has double-precision floats, Math.fround is used to cut off the mantissa to 32 bits for Float32 fields.

The mantissa of 32-bit floating-point numbers are 23 bits. I think the important aspect of Math.fround is that it "returns the nearest 32-bit single precision float representation of a number."

[hahnjo]

@undefined-panda can you please rebase to resolve the (trivial) conflict?

Proposal for the commit message:

Real32Quant is stored as integers on disk and then transformed back
to floating-point numbers using double-precision arithmetic. In the
ROOT (C++) implementation, the value is then cast according to the
requested field type, which can have either single-precision (float)
or double-precision.
Since JavaScript only has double-precision floats, Math.fround is used
to get the nearest 32-bit single-precision number.