Probe target compiler for opaque handle size

[krystophny]

Summary

• probe opaque derived-type handle size with --fortran-compiler, F90WRAP_FC, F90, or FC
• allow --sizeof-fortran-t for cross-compilers whose probe executable cannot run locally
• use one integer(c_int) element count in generated Fortran, Direct-C arrays, and generated Python array access
• add a Flang MRE and compiler-probe regression tests

The handle remains an opaque, byte-preserving container. The change does not interpret descriptor data or alter object ownership and lifetime.

This PR is stacked on #377 so its test fixture and manylinux_2_28 wheel fix are present. Once #377 merges, the remaining functional change is the issue #378 fix.

Fixes #378.

Verification

Test fails on main

$ f90wrap --direct-c -m example example.f90
$ flang-new -c example.f90 -o example.o
$ flang-new -c f90wrap_example.f90 -o f90wrap_example.o
error: Semantic errors in f90wrap_example.f90
./f90wrap_example.f90:44:5: error: Dimension 1 of left-hand side has extent 4, but right-hand side has extent 10
      this = transfer(this_ptr, this)
      ^^^^

Test passes after fix

$ make -C examples/issue378_flang_handle_size clean test F90=flang-new
flang-new -fPIC -c example.f90 -o example.o
flang-new -fPIC -c f90wrap_example.f90 -o f90wrap_example.o

The gfortran MRE also passes.

Regression suite

$ python -m pytest -q test/test_directc.py test/test_pywrapgen.py test/test_sizeof_fortran.py
36 passed in 0.26s

Wheel build, isolated installation, compiled-extension import, and both Meson example entries pass. --build --fortran-compiler flang-new produces objects identified as LLVM Flang 22.1.6.

RABE with Flang

Validated against RABE PR #76 with f90wrap built by gfortran and wrappers built by Flang:

[230/230] Linking Fortran shared module python/_rabe...so
100% tests passed, 0 tests failed out of 32
handle_len 10
b_mod 1.01
covariant (0.0, 8.0)

[GeorgGrassler]

@krystophny out of interest: this is a targeted fix of the --direct-c functionality, right? Or does it also affect the classic mode?

[krystophny]

It affects both modes. The handle size is resolved before the --direct-c branch and is passed into the classic Fortran-wrapper transformation/generator and the generated Python array-access code. In Direct-C mode, the same value additionally sizes the generated C handle buffers. So issue #378 exposed it with flang/Direct-C, but the fix also prevents classic wrappers from using a handle size derived from a different compiler. (If no target compiler or explicit size is supplied, it still falls back to the installed f90wrap value.)