Import Gmsh meshes at any order; finish tensor_dofmap 3D face gluing

Three related changes:

1. tensor_dofmap (src/TensorFEM.jl): implement the previously-unsupported case
   of shared 3D face-interior grids (d=3, k>=3). A face-interior node lies on a
   quad face shared between two hexes, which the two elements present in one of
   the face's 8 symmetries. New helper _tf_face_pos canonicalizes it: anchor at
   the minimum-id corner (unique for a valid face), measure the node from it
   along each axis, and order the two axes by the neighbour ids so axis 1 points
   toward the smaller-id neighbour. The key is then identical from both elements'
   framings, so the (k-1)^2 interior grid glues node-for-node. Subsumes the old
   k=2 special case; only interior grids on shared entities of dimension >= 3
   (possible only at d >= 4) remain unimplemented. tensor_dofmap now supports any
   k for d <= 3.

2. Gmsh import (ext/MultiGridBarrierGmshExt): resample element geometry at
   MultiGridBarrier's Chebyshev reference nodes via getJacobians, so quads and
   hexes import at ANY order (Gmsh's equispaced high-order nodes otherwise only
   match the Chebyshev reference through k=2). Physical-group membership uses a
   corner-subset test, exact at any order. With tensor_dofmap finished, the
   earlier 3D-hex k<=2 cap is removed.

3. CI (.github/workflows/CI.yml): after deploydocs, POST to the Pages builds API
   to force a rebuild -- GitHub's legacy Pages builder intermittently hangs and
   leaves the published docs stale.

Tests: transposed-2-hex tensor_dofmap oracle (shared face axes swapped) matches
coordinate-dedup exactly at k=2,3,4; Gmsh quads k=1..4, curved disk, hexes
k=1..3; full Pkg.test green, docs build clean.

Author: Sébastien Loisel

.github/workflows/CI.yml

@@ -98,3 +98,12 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           DOCUMENTER_KEY: ${{ secrets.DOCUMENTER_KEY }}
         run: xvfb-run --auto-servernum julia --project=docs --color=yes docs/make.jl
+      # deploydocs pushes the built site to the gh-pages branch, but GitHub's
+      # legacy Pages builder intermittently hangs (status "building", duration 0),
+      # leaving the published site stale. Force a fresh Pages build so it updates.
+      # Only on the events where deploydocs actually publishes (push to main / tags).
+      - name: Nudge GitHub Pages build
+        if: github.event_name == 'push' && (github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/tags/'))
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: gh api -X POST repos/${{ github.repository }}/pages/builds

docs/src/gmsh.md

@@ -13,9 +13,9 @@ automatically when both packages are imported and provides
 into a `Geometry` plus named node sets:
 
 - 3/6-node triangles → `fem2d_P1` / `fem2d_P2` (curved P2 edges supported),
-- 4/9-node quadrilaterals → tensor `fem2d` with `k = 1/2` (curved quads
-  supported; non-planar quad meshes become embedded surfaces),
-- 8/27-node hexahedra → `fem3d` with `k = 1/2`.
+- quadrilaterals → tensor `fem2d` of **any order** (curved; non-planar quad
+  meshes become embedded surfaces),
+- hexahedra → `fem3d` of **any order**.
 
 Gmsh **physical groups** come back as `(vertex, element)` node-pair lists — the
 same format as [`find_boundary`](@ref) — so named boundary parts plug directly
@@ -83,16 +83,21 @@ close()  # hide
 ```
 ![](gmsh_hole.svg)
 
-## Curved elements
+## Curved elements, any order
 
-Second-order meshes import as isoparametric elements: `setOrder(2)` before
-importing gives curved 6-node triangles (`fem2d_P2`) or curved 9-node quads
-(tensor `fem2d`, `k = 2`) whose boundary nodes lie on the true geometry. For
-all-quad meshes, set `Mesh.RecombineAll = 1` (with
+High-order meshes import as isoparametric elements: call `setOrder(k)` before
+importing. Triangles are P2 (`setOrder(2)`); quadrilaterals support **any order**
+— the geometry is resampled at MultiGridBarrier's Chebyshev reference nodes, so a
+curved `Q_k` element imports exactly regardless of `k`. Boundary nodes lie on the
+true geometry. For all-quad meshes set `Mesh.RecombineAll = 1` (with
 `Mesh.SubdivisionAlgorithm = 1` to guarantee no leftover triangles); for
 hexahedra use transfinite/swept volumes, or subdivide a tet mesh with
-`Mesh.SubdivisionAlgorithm = 2`. Orders ≥ 3, serendipity elements, and
-tetrahedra are rejected with actionable error messages.
+`Mesh.SubdivisionAlgorithm = 2`. Hexahedra also import at any order. Order-≥3
+triangles (MultiGridBarrier has only P1/P2 triangles), serendipity (incomplete
+high-order) elements, and tetrahedra are rejected with actionable messages.
+
+The example below imports a curved **fifth-order** disk and checks the disk
+area — the curved-boundary quadrature converges as the order rises:
 
 ```@example gmsh
 gmsh.initialize()
@@ -104,10 +109,10 @@ gmsh.option.setNumber("Mesh.MeshSizeMax", 0.25)
 gmsh.option.setNumber("Mesh.RecombineAll", 1)
 gmsh.option.setNumber("Mesh.SubdivisionAlgorithm", 1)
 gmsh.model.mesh.generate(2)
-gmsh.model.mesh.setOrder(2)          # curved Q2 quads
+gmsh.model.mesh.setOrder(5)          # curved Q5 quads
 gmd = gmsh_import()
 gmsh.finalize()
-sum(gmd.geometry.w) - π              # curved quadrature: area error of the disk
+gmd.geometry.discretization.k, sum(gmd.geometry.w) - π   # order, and disk-area error
 ```
 
 ## API reference