This guarantees floating-point -> text -> floating-point
roundtrip.

Those places, where the output precision was explicitly
set to 15-18 places, were replaced with max_digits10
precision.

The legacy reader is designed for unsigned type of material ids,
but current implementation works on signed integers.

This fixes a compilation warning.

Remove unused version of findElementOrError with iterators
and error string.

Significantly faster code, 3min to 30s for steady state cube 1e6 ctest.

Reading the VTU files can trigger floating point exceptions. Because
we are not debugging VTK (or other libraries) at this point, the
floating point exceptions are temporarily disabled and are restored
at the end of the scope.

See https://forum.hdfgroup.org/t/h5dwrite-fails-for-larger-data-and-or-smaller-chunks/11677

- Is currently not used in OGS.
- Fixes Linux ARM compilation.

Often used in sorting and the like algorithms on node and element vectors

and used new version for mat_ids