Add OpenMP hybrid parallelism to RRTMGP-LW block loop in LW_Driver

[tclune]

Motivation

Following the successful OpenMP parallelization of the RRTMGP-SW block loop in GEOS_SolarGridComp.F90 (PR #86, 1.5-1.75x speedup at 2 threads), this issue tracks the same work for the RRTMGP-LW block loop in GEOS_IrradGridComp.F90.

The target loop is do b = 1, nBlocks in LW_Driver (line 2778). Each iteration is fully independent -- it operates on a non-overlapping column slice colS:colE of the input/output arrays and uses thread-local scratch objects. RRTMGP is confirmed thread-safe by its developers.

Approach

Mirror the Solar approach exactly:

1. Step 1 -- Uniform per-block alloc/dealloc: replace the two-phase partial-block pattern with ceiling division so every iteration is structurally identical.
2. Step 2a -- Extract compute_lw_aer_optics
3. Step 2b -- Extract compute_lw_cloud_optics_mcica
4. Step 2c -- Extract compute_lw_gas_optics
5. Step 2d -- Extract compute_lw_rte
6. Step 3 -- Assemble PROCESS_RRTMGP_LW_BLOCK wrapper
7. Step 4 -- Add !$OMP PARALLEL DO SCHEDULE(DYNAMIC)

Each step is a separate commit. Steps 1-3 are pure refactoring (bit-identical results expected). Step 4 introduces threading (results may differ due to McICA stochasticity).

Key LW differences vs Solar

• Up to 4 rte_lw calls per block (clean clear-sky, clean all-sky, dirty clear-sky, dirty all-sky)
• Both ty_fluxes_broadband and ty_fluxes_byband flux structs used
• Surface Jacobians (flux_up_Jac, dfupdts_*)
• dirty_optical_props is a per-block copy of clean_optical_props + aerosol increment
• Planck source function (ty_source_func_lw sources) is a per-block object
• PRNG seeding uses direct global grid math (jBeg, iBeg, IM_World) rather than lookup arrays

Branch

feature/tclune/lw_openmp off feature/sdrabenh/gcm_v12