Make the surface samplers seedable (closes #54)

[gattia]

Closes #54.

Mesh.rand_pts_around_surface() had two independent random draws that a caller could not
reach, so identical inputs produced a different point cloud on every call. Both are now
seeded through a single optional seed argument.

Seeding only one of the two would have left the function random, which is the main reason
this was easy to miss — so every test here checks the whole call rather than one draw.

The change

• rand_sample_pts_mesh(..., seed=None) passes random_seed to both
  pcu.sample_mesh_random and pcu.sample_mesh_poisson_disk.
• Mesh.rand_surface_pts(..., seed=None) forwards it.
• Mesh.rand_pts_around_surface(..., seed=None) uses np.random.default_rng(seed) for the
  offsets instead of np.random.default_rng(), and forwards the seed to the surface draw.

pcu_random_seed()

pcu documents random_seed=0 as "use the current time", so a user's seed cannot be
handed over directly: seed=0 would silently mean "unseeded", and 0 is the first seed most
people try. pcu_random_seed() maps None to pcu's own 0 sentinel (preserving the
existing default) and any integer seed to a deterministic non-zero 32-bit value.

Backwards compatibility

seed=None is the default and behaves exactly as before — unseeded, and unaffected by
np.random.seed(), which never reached either draw. There is a test asserting that,
because it is the property a patch release must not break.

Verified, not assumed

Case	Result
same seed twice	identical
seed=0 twice	identical (the pcu trap)
different seeds	differ
seed=None twice	still differ — unchanged default
distribution="laplace"	seeded too
rand_sample_pts_mesh, both random and bluenoise	seeded

14 new tests in testing/mesh/meshTools/rand_sample_pts_mesh_seed_test.py.

One thing this does NOT fix

rand_pts_around_surface(surface_method="bluenoise") — the default surface_method —
is broken independently of seeding and remains broken:

ValueError: operands could not be broadcast together with shapes (1436,3) (1000,3)

pcu.sample_mesh_poisson_disk returns approximately num_samples points (it took 1450
for a requested 1000), and base_pts + rand_pts then broadcasts mismatched shapes. This
reproduces on main at 0.1.20, so it predates this branch. It is left alone deliberately:
it is a different defect with a design question attached (truncate? oversample and trim?
raise?), and bundling it into a seeding patch release would hide it.

Consequence for this PR's coverage: bluenoise seeding is verified at the
rand_sample_pts_mesh level, which is where the seed lands, but not end-to-end through
rand_pts_around_surface, because that path raises before it can be checked. The tests say
so at the point where it matters.

Version bumped 0.1.20 → 0.1.21.