Three read-only substitutions for repeated star/point-location queries

[IasonManolas]

Summary

Three independent optimizations, grouped together because each is a small, self-contained
substitution of a cheaper read for a query that a nearby invariant already answers:

• Read the subdomain of an edge midpoint off the edge. midpoint_with_info() located a
  non-boundary edge's midpoint with subdomain_index_at_point_3(), a full point-location query,
  purely to learn its subdomain. Every cell incident to a non-boundary edge shares one subdomain,
  since is_boundary() is exactly the test that rules out a facet of the complex or a subdomain
  change around the edge. The midpoint lies on the edge, so a locate can only return one of those
  incident cells, and e.first already is one. Reading e.first->subdomain_index() gives the
  same answer without the query. The locate is kept as a CGAL_expensive_assertion.
• Look for the surface patch without collecting the star. surface_patch_index() collected a
  vertex's entire incident-facet star into a small_vector, then scanned it for the first facet
  belonging to the complex. Streaming through a function output iterator instead removes the
  collection and stops at the first match, same as the scan already did.
• Examine each neighbour once when validating lengths. are_edge_lengths_valid() visits the
  incident edges of both endpoints of a collapsed edge, so a vertex adjacent to both is visited
  twice; the squared length was recomputed and re-queried from the sizing field on the repeat
  even though the loop's own test reads nothing but the vertex handle. Tracking already-examined
  vertices in a flat_set and skipping repeats removes the redundant sizing-field query.

Testing

Verdict equivalence. All three changes are read-only substitutions or streaming equivalents
over predicates already determined by the triangulation's invariants; none change which edge,
vertex, or collapse is selected, or in what order. Full test suite (all executables in
test/Tetrahedral_remeshing) passes.

Byte-identical output, against plain cgal/main:

mesh	tets	byte-identical
bear	~40K	YES
bunny00	~225K	YES
101556.mesh (Thingi10K)	3,521,237	YES

Performance. perf stat -e instructions, Linux, Release, sequential, against plain
cgal/main.

mesh	iterations	cgal/main instr	this MR	Δ instr
bear	10	67.04 G	58.35 G	-12.97%
bunny00	10	148.80 G	130.91 G	-12.02%
101556.mesh	3	2441.29 G	2263.33 G	-7.29%

Release Management

• Affected package(s): Tetrahedral_remeshing
• License and copyright ownership: unchanged