Enable -O0 and runtime diagnostics in DEBUG build for ICON

[s-poll]

Summary

CMAKE_BUILD_TYPE=DEBUG currently only adds -g to the ICON build. Under Intel, no -O level is set explicitly, so the compiler defaults to -O2 — the "debug" build is actually optimized and gets no runtime checking at all. This made a recent segfault in ecRad take roughly a day to isolate; -check bounds alone would have pointed at it in minutes.

This PR makes DEBUG a real debug build: unoptimized and instrumented with runtime diagnostics, for both compilers.

Solves #137

Changes

In cmake/BuildICON.cmake, the DEBUG branch now sets:

• -O0 (explicit, both compilers) — closes the implicit -O2 gap under Intel
• Bounds checking: -check bounds (Intel) / -fcheck=bounds (GNU)
• Uninitialized-variable detection: -check uninit (Intel); on GNU there's no direct runtime-uninit-check equivalent, so this is approximated via NaN-poisoning + FP trapping below
• FP exception trapping: -fpe0 (Intel) / -ffpe-trap=invalid,zero,overflow (GNU)
• NaN-initialization of reals: -init=snan (Intel) / -finit-real=snan (GNU)

Flags are gated on CMAKE_Fortran_COMPILER_ID since the Intel spellings (-check, -init=) aren't recognized by gfortran.

RELEASE (and the unknown-build-type fallback) are unchanged.

Trade-off

These diagnostics cost wall-clock time - bounds checking in particular adds real per-access overhead in hot loops, on top of the much larger cost of dropping to -O0. That's expected and acceptable for a debug build; -fpe0/-ffpe-trap and the NaN-init are effectively free (one-time trap-mask/init cost, not per-operation).