Replace three enclosing-scope fypp macros with explicit procedures

[sbryngelson]

Turns the fypp macros that expand into a caller's scope into ordinary procedures with explicit
interfaces. First three of the nine in MFlowCode#1769; six of the twenty-six call sites.

A zero-argument macro reads and assigns the local variables of whatever routine expands it, so a
call site says nothing about what it touches and neither the compiler nor a reader can check the
coupling. MFlowCode#1769 has the worked example: roe_avg accumulates vel_avg_rms over every velocity
component and overwrites it with the first component six lines later, which is a visible double
assignment in a subroutine and invisible inside a 45-line macro.

In this PR

macro	sites	becomes
hll_flux_component	3	f_hll_flux(S_L, S_R, F_L_i, F_R_i, U_L_i, U_R_i)
compute_capillary_stress_tensor	3	s_compute_capillary_stress_tensor(sigma, w1, w2, w3, normW, Omega)
compute_axis_inv_re	4	s_compute_axis_inv_re(grad_x_vf, grad_y_vf, grad_z_vf, alpha_visc, j, k, l, Re_visc)

include/inline_capillary.fpp held nothing but the second macro and is deleted with it.

Two details worth review rather than assumption:

Omega is intent(inout), not intent(out). The macro assigned only the entries a given
dimensionality defines, leaving the rest as the caller had them; intent(out) would make them
undefined in 1D and 2D.

gamma_dot_c was live only inside compute_axis_inv_re, so it becomes a local of the new routine
and leaves the caller's declarations and all four private() clauses.

Cost

compute_axis_inv_re takes three scalar_field array dummies on a [seq] device routine, which is
the shape MFlowCode#1714 showed collapsing register promotion across a whole loop body for 20%. It does not
reproduce. The distinction appears to be real: MFlowCode#1714 constructed a derived type and took its
address, while these arrays already exist and are passed by reference.

viscous_weno5_sgb_acoustic, MI210, AFAR amdflang, OpenMP offload, three runs each on one node:

	mean ns/gp/eq/rhs
before	3.4471
after	3.4297

-0.50%, with every after run below every before run - consistent with the smaller private()
clauses rather than with luck.

The other two macros have no benchmark that reaches them: no benchmark case enables surface tension,
and f_hll_flux lives in the hypoelastic HLLD solver while the only hypoelastic benchmark uses HLL.
For those the evidence is the offload image: identical scratch, VGPR and AGPR across all 471 kernels,
which is the mechanism a non-inlined call would have to show up in.

One kernel does move: m_riemann_solver_hlld gains 16 B of scratch on 21440, in a module none of
these changes can reach. That is the whole-image codegen variance of MFlowCode#1759. Every kernel actually
touched here is unchanged.

Tests

Hypoelasticity 59, Viscous 63, the four capillary and surface_tension cases - 126 passed, 0 failed.
No goldens regenerated.

Remaining

Six macros, twenty call sites: compute_low_Mach_correction (8, and it silently overwrites
vel_L(dir_idx(1)) and vel_R(dir_idx(1))), compute_elastic_wave_speeds_lr (2),
compute_hypo_elastic_energy (1), and arithmetic_avg/roe_avg/compute_average_state (5), which
are widest and should go last - the chemistry path there produces six further outputs and wants its
own routine.

Whether roe_avg's dead loop should lose the loop or the overwrite is a physics question, not a
refactoring one, and belongs in its own change.

Stacked on MFlowCode#1762 because m_riemann_solver_hypo_hlld.fpp is not on master yet.

[sbryngelson]

Folded into MFlowCode#1762 — these commits are now on that branch, so this closed itself when it was pushed. The macro work was always stacked on fix/speed-of-sound-from-state, so a separate fork PR only added indirection.

Landed there: 6 of 9 inline_riemann/inline_capillary macros converted to procedures (21 of 26 call sites), plus a Roe-average bug the conversion turned up. compute_average_state, roe_avg and arithmetic_avg remain; landing them deletes inline_riemann.fpp and closes MFlowCode#1769.