Tetrahedral_remeshing: replace hash containers with linear scans over small, bounded sets

[IasonManolas]

Summary

Four independent optimizations, grouped together because each replaces a hash-based container
with a direct lookup or a linear scan, in places where the container never holds more than a
handful of elements, so hashing costs more than it saves:

• Ask the triangulation which facets bound the cell set. is_cells_set_manifold() counted,
  for each facet of a candidate cell set, how many times each vertex triple appears, via a hashed
  vertex-triple-to-count map, to find facets shared by exactly one cell (the boundary of the
  set). The triangulation already knows which facets border a given cell set: a direct
  cell-neighbor lookup replaces the hashed triple counting.
• Build the collapse simulation without hash containers. CollapseTriangulation's
  constructor numbered the star's roughly twenty vertices via an unordered_set plus an
  unordered_map, hashing a compact-container iterator for each. A small_vector of pairs with
  linear scan replaces both, since a star this small is cheaper to scan than to hash.
• Search the collapse work list by hash, not in order. The collapse work list is a bimap
  whose element side was set_of<Edge, Compare_edges>, an ordered container, so every membership
  test (the priority update or erase that follows each collapse) cost O(log n) comparisons, each
  building and comparing canonical vertex pairs. That side is never iterated in order; only the
  priority side is, and it is untouched. Switched to unordered_set_of with matching
  Hash_edges/Equal_edges.
• Count the indices around a simplex without a hash set. nb_incident_subdomains() and
  nb_incident_surface_patches() filled an unordered_set with every index seen around a vertex
  or edge just to read its size back out, for a count that is almost always 1 or 2. Replaced with
  an inline small_vector and linear dedup.

Testing

Verdict equivalence. All four changes are alternative implementations of the same predicate
or count; none change which candidate is evaluated, 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, measured with
PR #9587 applied as the base, since that PR's decisive-accept fast path already skips
CollapseTriangulation for most candidates and changes how often the second optimization above
gets exercised. Byte-identical to that base on all three meshes.

mesh	iterations	base (#9587) instr	this MR	Δ instr
bear	10	50.40 G	47.40 G	-5.95%
bunny00	10	123.35 G	117.05 G	-5.11%
101556.mesh	3	1775.84 G	1669.82 G	-5.97%

Release Management

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