Treat JAX and NumPy arrays as interchangeable when deserialising

[JOhnsonKC201]

Fixes #861.

Problem

Deserialising into a filter_eval_shape skeleton fails if the filter_spec
puts the result on the host:

like = eqx.filter_eval_shape(eqx.nn.Linear, 2, 2, key=key)
eqx.tree_deserialise_leaves(
    path, like,
    filter_spec=lambda f, x: jax.device_get(eqx.default_deserialise_filter_spec(f, x)),
)
# RuntimeError: Deserialised leaf at path '.weight' has changed type from
# <class 'jax.ShapeDtypeStruct'> in `like` to <class 'numpy.ndarray'> on disk.

Shape and dtype match. Only the array kind differs.

Cause

_assert_same_impl substitutes the concrete array type for a
ShapeDtypeStruct in like, then compares with typenew is not typeold. That
identity check leaves no room for a filter_spec to hand back a NumPy array
instead of a JAX one.

Fix

Compare array leaves by array-ness instead of exact type, which is what you
suggested in the issue: "Maybe we just consider all kinds of JAX and NumPy array
interchangeable?".

Every non-array leaf keeps the strict identity check, and the shape and dtype
checks immediately below are untouched, so real mismatches still raise. I
verified both: a shape mismatch still errors, and the behaviour for a
non-array like leaf is byte-for-byte the same as on main.

Test plan

• New test_eval_shape_deserialise_to_host in tests/test_serialisation.py.
  It fails on main with the RuntimeError above and passes with this change.
• Full suite: 472 passed, 4 skipped (baseline on a clean checkout was 471
  passed, 4 skipped, so the only difference is the added test).
• ruff check and ruff format --check clean on both files.