Initialize model _parameters array at construction (fix flaky Gaussian1D pickling) - #20060

What this fixes

Fixes #18598.

A freshly constructed Model allocated its _parameters array with np.empty()
in _initialize_slices() and only populated it lazily, on the first access of the
parameters property. Until that first access, _parameters held uninitialized
memory.

That is harmless for the normal fitting workflow (which always touches
.parameters), but it makes any operation that reads the raw _parameters array
directly non-deterministic. The clearest victim is pickle-recovery testing:
astropy.tests.helper.check_pickling_recovery() compares the original and
unpickled objects attribute-by-attribute with strict equality. When the
uninitialized memory happened to contain a NaN, the comparison
original._parameters == unpickled._parameters produced False (since
NaN != NaN), raising AssertionError: Value of _parameters changed by pickling.

This is exactly the intermittent failure that led to Gaussian1DKernel being
disabled from convolution/tests/test_pickle.py in #18596 — Gaussian1DKernel
wraps a Gaussian1D, so the modeling-level flakiness surfaced there. As #18598
notes, fixing Gaussian1D fixes the kernel automatically.

The fix

Populate _parameters from the actual parameter values immediately after it is
allocated, so a pristine model never holds uninitialized memory:

self._parameters = np.empty(total_size, dtype=np.float64)
self._parameters_to_array()   # fill with real values right away

_parameters_to_array() only needs the _param_metrics slices (already built
earlier in _initialize_slices) and the parameter values, so it is safe to call
here. After the change, a fresh Gaussian1D() has _parameters == [1., 0., 1.]
(its real defaults) instead of garbage.

The disabled Gaussian1DKernel case in convolution/tests/test_pickle.py is
re-enabled (this is the Revert #18596 item on the issue checklist).

How it was tested

Reproduced the failure first (RED). With a fresh model whose uninitialized
_parameters was forced to contain NaN, check_pickling_recovery() failed with
the reported signature:

check_pickling_recovery: FAILED (RED) -> Value of _parameters changed by pickling

Reverting the one-line fix in the working tree and running the new
test_fresh_model_parameters_are_initialized shows the raw array actually holding
uninitialized garbage rather than the parameter values:

ACTUAL:  array([ 1.797693e+308, -1.797693e+308, 2.0, 2.22e-16, 2.22e-308, 4.94e-324])
DESIRED: array([1., 0., 0., 1., 1., 0.])

With the fix applied (GREEN), the full suites pass:

• pytest astropy/modeling/tests/ → 1690 passed, 1118 skipped (scipy-gated)
• pytest astropy/convolution/tests/ → 1560 passed, 36 skipped, 8 xfailed
• combined modeling + convolution → 3250 passed, 0 failed

ruff check and ruff format --check (pinned v0.15.20) are clean on all three
changed files.

Two regression tests are added in modeling/tests/test_pickle.py:
test_fresh_model_parameters_are_initialized (asserts a pristine model's raw
_parameters matches its real values and contains no NaN) and
test_gaussian1d_strict_pickling_recovery (exercises the strict
check_pickling_recovery path across all pickle protocols).