dirichlet_from_buffers is GPU_DECLAREd and GPU_UPDATEd from different modules; a missed update silently degrades Dirichlet BCs

[sbryngelson]

dirichlet_from_buffers is declared and device-mapped in one module but assigned and updated from another:

• src/common/m_boundary_primitives.fpp:21-22 — logical :: dirichlet_from_buffers = .false. + GPU_DECLARE(create=...)
• src/common/m_boundary_common.fpp:42-44 — assigned, then GPU_UPDATE(device='[dirichlet_from_buffers]')
• src/common/m_boundary_primitives.fpp:886 — read on device: if (.not. dirichlet_from_buffers) then

This is the same cross-module declare/update idiom as Re_idx, so it is not novel. The concern is the failure mode: if the device update ever fails to take, the flag reads .false. on device and Dirichlet BCs silently degrade to zero-gradient — wrong boundary physics with no error, no NaN, and no crash.

Not a known-failing case today; filing so the fragility is tracked rather than rediscovered from a wrong answer.

Suggested check: confirm on Cray CCE (where declare/update attachment has bitten us before) that the device value actually reflects the host assignment at the read site, e.g. a one-off assertion or a targeted case with use_dirichlet_buffers = .true.