[BUG] get_params/set_params/clone are silently broken for default-constructed estimators

[ChrisW09]

Describe the bug
For default-constructed estimators (no explicit config objects), the sklearn parameter API is silently broken (deeptab/models/base.py:258-358):

• get_params() returns {} — it drops random_state entirely (the legacy branch only returns _config_kwargs + preprocessor kwargs), and observability_config is missing from both branches.
• clone(model) consequently loses random_state (and any other constructor state).
• set_params(layer_sizes=[8]) only updates the _config_kwargs shadow dict, which is never used to rebuild self.config (verified by grep: _config_kwargs is only read back by get_params and serialization). The model still trains with the old architecture — while get_params() echoes the new value back, so the API actively lies.
• Unknown/invalid parameter names are silently accepted in both branches; sklearn requires a ValueError.

Net effect: GridSearchCV/RandomizedSearchCV over a default-constructed deeptab estimator runs to completion while never changing a single hyperparameter.

To Reproduce

from sklearn.base import clone
from deeptab.models import MLPClassifier

m = MLPClassifier(random_state=42)
print(m.get_params())                  # {}  (random_state gone)
print(clone(m).random_state)           # None

m.set_params(layer_sizes=[8])
print(m.config.layer_sizes)            # [256, 128, 32]  -- unchanged!
print(m.get_params())                  # {'layer_sizes': [8]}  -- but echoed back

m.set_params(totally_bogus_param=123)  # silently accepted, no ValueError

Expected behavior
get_params/set_params/clone must round-trip all constructor state and actually affect the next fit; invalid names must raise ValueError. The split-config path (explicit model_config=/trainer_config=) behaves much better — the legacy flat-kwargs path should either be wired to the real config or removed.

Screenshots
n/a

Desktop (please complete the following information):

• OS: macOS (Darwin 25.5.0, arm64)
• Python version: 3.11.15
• deeptab Version: 2.0.0 (main @ 4e6a359)

Additional context
The sklearn contract tests in tests/test_sklearn_contract.py pass because they parametrize estimators with explicit config objects, which takes the (healthier) split-config branch — the default-constructor path is never exercised. Related: __getstate__ (base.py:371) nulls state["task_model"] but the attribute is _task_model, so pickling serializes the full Lightning module despite the documented intent in _mixins/serialization.py (filed separately in the robustness umbrella issue).