Refactor case BC hooks: rename to define_BC/apply_BC and hoist BC fields to base

[kaanolgu]

• Rename boundary_conditions -> define_BC and pre_correction -> apply_BC across base and child cases.
• Hoist 12 BC field pointers (bc_{start,end}_{u,v,w}_{x,y}) onto base_case_t; cases use only what they need.
• Add field_set_face_from_field backend hook (CUDA + OMP) for spatially-varying face BCs.
• Channel: y-wall BCs now use persistent device fields refreshed with noise per substep, applied via the new hook.
• Cylinder: inflow profile now built field-side using new inlet_noise(3) config (replaces bc_start_u/v/w scalars).
• Cylinder: compute_outflow_params moved from apply_BC into define_BC; results stored on self%out_vel / self%flow_rate_diff (drops _cached suffix).
• Behavioural note: outflow params sampled one substep earlier than before.
• Breaking change: existing cylinder_nml files must replace bc_start_u/v/w with inlet_noise.

[ia267]

I think this should be:

Suggested change

	y_block = (k - 1)/dims(3) + 1
	y_block = mod(k - 1, n_y_blocks) + 1

OMP reordering uses get_index_dir (line 54), which maps a Cartesian (y, z) position to the third DIR_X index as:

dir_k = n_y_blocks*(z-1) + y_block

Therefore, y-block is the fast-varying component. For n_y_blocks = 3 and nz = 2, increasing k from 1 to 6 gives y-blocks 1, 2, 3, 1, 2, 3. The current expression, (b - 1)/nz + 1 produces 1, 1, 2, 2, 3, 3 (opposite flattening order). For CUDA it calls reorder_cuda (doesn't call get_index_dir). The current expression assumes CUDA's different layout.

Also, I think in field_set_face_omp (line 848), the Y_FACE block also uses the same incorrect CUDA-style indexing and should be changed to:

case (Y_FACE)
  i_mod = mod(dims(2) - 1, SZ) + 1
  n_y_blocks = (dims(2) - 1)/SZ + 1

  !$omp parallel do private(k_start, k_end)
  do z = 1, dims(3)
    k_start = 1 + (z - 1)*n_y_blocks
    k_end = n_y_blocks + (z - 1)*n_y_blocks

    do j = 1, dims(1)
      f%data(1, j, k_start) = c_start
      f%data(i_mod, j, k_end) = c_end
    end do
  end do
  !$omp end parallel do

This follows the OMP's DIR_X layout.

[ia267]

The channel case calls this routine with Y_FACE - we will need to implement Y_FACE path using OMP ordering:

n_mod = mod(dims(2) - 1, SZ) + 1
n_y_blocks = (dims(2) - 1)/SZ + 1

!$omp parallel do private(k_start, k_end)
do z = 1, dims(3)
  k_start = 1 + (z - 1)*n_y_blocks
  k_end = n_y_blocks + (z - 1)*n_y_blocks

  do j = 1, dims(1)
    f%data(1, j, k_start) = f_start%data(1, j, k_start)
    f%data(n_mod, j, k_end) = f_start%data(n_mod, j, k_end)
  end do
end do
!$omp end parallel do

[ia267]

The OMP boundary path incorrectly assumes CUDA’s z-fast block ordering. OMP uses y-block-fast ordering through get_index_dir, so the y_block calculation and Y-face indices must be corrected accordingly. Additionally, field_set_face_from_field_omp must support Y_FACE, since the channel case calls that path.

[ia267]

Approved - please fix the formatting issues before merging.

[kaanolgu]

@x3d2bot fix style

[github-actions]

On it -- running fprettify to fix style issues.