fix(conversions): cast geom_type to int in merge_geoms enum check

[Guozhongyuan]

Problem

merge_geoms() classifies geoms with:

if mj_model.geom_type[geom_id] in (mjtGeom.mjGEOM_MESH, mjtGeom.mjGEOM_SDF):

MjModel.geom_type[i] is a NumPy scalar (np.int32). Tuple membership evaluates
the element on the left (enum_member == np_scalar), and MuJoCo's pybind11 enum
returns False when compared against a NumPy scalar from that side — even though
the identical comparison with the NumPy scalar on the left is True:

>>> import mujoco, numpy as np
>>> E = mujoco.mjtGeom
>>> np.int32(7) == E.mjGEOM_MESH              # numpy scalar on the left
True
>>> E.mjGEOM_MESH == np.int32(7)              # enum on the left — what `in` does
False
>>> np.int32(7) in (E.mjGEOM_MESH, E.mjGEOM_SDF)
False

So the membership test silently fails for every geom type. Mesh and SDF geoms
fall into the create_primitive_mesh() branch and raise:

ValueError: Unsupported shape type: 7

Primitive geoms keep working only because the primitive fallback happens to
reconstruct the same shape. Any model with a mesh geom among its fixed-body
geometries — i.e. almost every real model — crashes when its static geoms are
merged. This also means the SDF routing added in #31 never actually triggers.

merge_geoms_hull() is unaffected because it already compares via int(...),
which is also why this slipped through CI: the existing test_merge_geoms only
merges primitive geoms.

Fix

Cast both sides to int() before the membership test, matching the existing
idiom in merge_geoms_hull(). One-line change plus a regression test.

Testing

• Added test_merge_geoms_with_mesh_geom, which merges a mjGEOM_MESH geom:
  raises ValueError: Unsupported shape type: 7 before the fix, passes after
  (verified red/green).
• Full suite: 75 passed.

Reproduced on mujoco 3.14.0, numpy 2.5.3, Python 3.12. The regression was
introduced by #31 (post-v0.0.14, main only), which is why released wheels are
unaffected.