Fixed per-thread array bounds on all GPU compilers, with a generated state-vector layout #2081

Summary

Per-thread arrays in GPU kernels now get compile-time extents in every GPU simulation build, from one place, and case optimization supplies exact extents including sys_size.

Runtime-sized private arrays (dimension(num_fluids), dimension(sys_size)) cannot live in registers and spill to scratch. The USING_AMD guards fixed this for amdflang only; profiling on AFAR 24.3 (rocprofv3) showed WENO 11x and HLLC 4.9x slower without them. #1938 turned the same fixed bounds on for every compiler, and CI showed CCE 12-36% and NVHPC OpenMP 9-62% faster (AMD unchanged), but GNU CPU up to 39% slower. So this PR applies them to GPU builds only.

Changes (four commits)

1. BOUND() replaces the USING_AMD guards. The 83 copy-pasted #:if [not MFC_CASE_OPTIMIZATION and] USING_AMD blocks become single declarations such as dimension(${BOUND('num_fluids')}$). BOUND returns the fixed maximum from one table (num_fluids 3, nb 3, sys_size 70 or 10 + species, WENO and QBMM extents, and so on) or the runtime extent. A converter checked every existing literal against that table. MFC_FIXED_BOUNDS (CMake, default ON) is decided per target: only the offloaded simulation uses fixed bounds, so pre_process/post_process callers always match the routines they call. s_check_amd becomes s_check_fixed_bounds, and the sys_size limit is checked right after the layout is computed (the old input-time check ran before sys_size existed).
2. The state-vector layout is generated from one table. toolchain/mfc/params/eqn_layout.py lists each model's fields with their size and condition as small Python expressions. The params codegen turns them into generated_eqn_idx.fpp, replacing about 140 hand-written lines of s_initialize_eqn_idx. Checked against master's routine over all 7,962,624 combinations of model, the 13 layout flags, 1D/multi-D, and the counts: every eqn_idx field and sys_size matches.
3. Case optimization bakes in exact extents. It evaluates the same table for the case and writes CASE_OPT_SIZES (sys_size and the hypoelastic stress count) into case.fpp, so BOUND gives exact compile-time extents under case optimization on every compiler. The sys_size variable stays a runtime variable, so nothing that declares or updates it on the device changes. Simulation checks at startup that the runtime layout matches the baked-in values. Since CASE_OPT_SIZES is in case.fpp, it is part of the build fingerprint.
4. Fixed bounds are on for every GPU compiler (NVHPC and CCE, OpenACC and OpenMP); CPU builds keep runtime extents. Docs describe BOUND().

Behavior change

GPU simulation builds without --case-optimization now accept at most 3 fluids, nb <= 3, and sys_size <= 70 (or 10 + species) on every GPU compiler, as amdflang already did; larger cases rebuild with --case-optimization, and the error message says so. CPU builds and pre_process/post_process are unaffected.

Testing

• Generated code: for amdflang simulation, and for gfortran simulation/pre_process/post_process, the preprocessed Fortran matches master apart from the checker rename and messages.
• HPCFund MI210, AFAR 24.3, --gpu mp, release: full test suite, 718/718 passed.
• gfortran CPU: full test suite, 718/718 passed.
• Case-optimized 5eq_rk3_weno3_hllc on GPU: runs with CASE_OPT_SIZES = {'sys_size': 8} baked in (3D, two fluids), HLLC and CBC arrays sized exactly 8, startup check passes; grind 1.93 against 2.50 without case optimization (--mem 2).
• AMD benchmark against master, same node (--mem 2): geometric-mean ratio 1.000, every case within ±2%. AMD already had fixed bounds, so no change is expected; the gains are on NVHPC and CCE.
• NVHPC and CCE: the CI Test and Bench jobs on this PR.

Status: a fix/implementation for this is already written, in a pull request that has been closed for queue management.

• Closed PR: Fixed per-thread array bounds on all GPU compilers, with a generated state-vector layout #1941
• Branch: fixed-bounds on the sbryngelson fork, at 2dcb6ed3c1 (branch retained, not deleted)
• Size: +954/-1226 across 39 files, 13 commits
• Where it sat: MERGEABLE; CI 19 checks queued (master was merged in today, so CI had not reported yet)
• Reopen with: gh pr reopen 1941 -R MFlowCode/MFC

It was merge-clean and had been brought up to date with master, so resuming should be straightforward.

Tracked with pr-to-do. Filed with Claude Code.