Use linear scan for flexcomp unused-point reindexing

[teerthsharma]

Summary

• replace flexcomp's per-unused-point suffix updates with an equivalent prefix count
• preserve reindex behavior for leading, interleaved, and trailing unused points
• add an end-to-end regression covering elements, texcoords, bodies, positions, pins, and flex vertices

Fixes #3422.

Performance

The old loop performs Theta(P * U) updates for P input points and U
unused points, which is Theta(P^2) in the worst case. The replacement is one
Theta(P) scan with constant additional state.

For a direct tetrahedral flexcomp with four referenced vertices and increasing
unused input points, median compile time at 32,000 points fell from 51.218 ms to
7.201 ms (7.11x faster, 85.9% less wall time). An all-points-used control found
no material regression. The temporary benchmark used to establish this result
is not included in the patch.

Validation

• Ubuntu 24.04, Clang 18.1.3, RelWithDebInfo + AddressSanitizer and leak
  detection: all 60 UserFlexTest cases passed
• Ubuntu 24.04, GCC 14.2, Release: full serial CTest suite passed
  1,926/1,926 tests (one declared locale-dependent test skipped)