Move particle-cloud packing out of the solver into pre_process

[sbryngelson]

Summary

Particle-cloud generation (src/simulation/m_particle_cloud.fpp) is geometry preprocessing, not flow solving — rejection sampling, triangular/FCC lattices, hemisphere-shell CDF sampling, a spatial hash grid, and periodic minimum-image wrapping have nothing to do with the solver. It currently lives in the simulation startup path only because IBM marker computation (m_ibm.fpp, s_ibm_setup) is simulation-side and the packer was placed next to it.

Proposal: expand the compact particle_cloud spec in pre_process, write the resulting IB-patch list to a file, and have the solver just read it. The solver becomes geometry-agnostic about beds and simulation/ sheds the whole subsystem.

This is a follow-up to the review on #1667 — not part of it. #1667 is a feature PR; this is a cross-executable refactor and should land after it merges, so we relocate a finished subsystem rather than a moving target.

Boundary

pre_process expands each particle_cloud into a global IB-patch list and writes restart_data/particle_cloud_ibs.dat; the solver reads that list at fresh start, then reduces to each rank's neighborhood (unchanged). ib_patch_parameters already lives in src/common/m_derived_types.fpp, so both stages share the type.

What moves

File	Change	~LOC
src/simulation/m_particle_cloud.fpp	deleted	-556
src/simulation/m_start_up.fpp	swap s_generate_particle_clouds(...) -> s_read_particle_cloud_ibs(...)	+30 / -5
src/pre_process/m_particle_cloud.fpp	new — the packers, minus the neighborhood filter (pre_process places the global set)	~+500
src/pre_process/m_start_up.fpp	run packer after grid/IC, write patch file	~+40
src/pre_process/m_global_parameters.fpp, m_mpi_proxy.fpp	add particle_cloud namelist awareness + broadcast (pre_process has none today)	~+20
src/common/	shared read/write helper for the patch file	~+40

Net: simulation/ loses ~525 lines (556-line subsystem -> ~30-line reader).

Design decisions

1. Reduction stays solver-side. f_neighborhood_ranks_own_location / s_reduce_ib_patch_array depend on the simulation's domain decomposition, so pre_process writes the global list and the solver reduces on read. The lattice packer's inline neighborhood filter disappears (it packs global now) — a simplification.
2. Scaling win, not loss. Today every rank redundantly recomputes the full deterministic placement; under the file boundary it is computed once. The list is bounded by num_ib_patches_max_namelist (54000) -> tiny file, so the old "avoid broadcasting N patches" concern goes away.
3. Restart path. On restart (t_step_start > 0) the solver already skips generation and reads ib_state; keep that branch. The .dat read fires only on fresh start (mirrors the current if/else in m_start_up.fpp:903-913).
4. RNG / goldens. Packing runs in a different binary, so regenerate the 6 particle-cloud goldens rather than assume a bit-identical xorshift stream. The ib_state geometric verifier in test.py is unchanged (reads solver output either way).
5. Validation. case_validator.py is untouched — it validates the compact user spec. This is the payoff of the input-only/runtime split from Move input-only checks to case_validator, delete dead params_tests, enforce the split #1717.

Test plan

• Build all three targets.
• Regenerate the 6 particle-cloud goldens.
• Add a ppn=2 case (classic "did the broadcast/decomposition survive" catch, per .claude/rules/common-pitfalls.md).
• Keep the geometric ib_state verifier.

Effort

Medium — ~80% mechanical relocation, ~20% genuinely new (file writer/reader + pre_process param plumbing). The real risk surface is parallel correctness, not the algorithms.