Fix shared-parameter checkpoints and check shapes on load

[cemde]

Bug

Two defects in _serialisation.py.

save_params deduplicates shared parameters by writing the value under the first path it meets and a literal None under every later one. np.savez_compressed stores that None as a 0-d dtype=object array. load_params then calls np.load, which refuses object arrays unless allow_pickle=True.

Separately, load_params assigned whatever the file held without comparing shapes.

Impact

Every checkpoint of a model using pxnn.shared is unrecoverable:

ValueError: Object arrays cannot be loaded when allow_pickle=False

The loader's is not None guard was dead code, since np.load yields a 0-d object array and never None, so even with allow_pickle=True the shared parameter would have been assigned garbage.

Without a shape check, a Linear(5, 7) checkpoint loaded into a Linear(2, 3) leaves the weight at (7, 5), surfacing much later as an inscrutable broadcasting error inside a training step.

Fix

The sentinel the writer and reader had to agree on turned out to be no key at all. save_params drops the else branch, so duplicates are simply not written; load_params mirrors it with the same _cache() idiom, skipping the same paths. No object array is ever created, so nothing needs allow_pickle.

Plus a shape comparison raising ValueError before assignment.

Compatibility

Old checkpoints still load, including shared ones that previously could not. np.load on an .npz is lazy and the loader no longer visits the duplicate paths, so the stale dtype=object entries are never read.

The shape check runs only when the target parameter already holds a value. A freshly constructed Vode has h set to None, and model = Model(); load_params(model, ...) is the documented workflow, so an unconditional _param.shape raised AttributeError on the normal path. Found in review; regression test added, since every sibling test seeded its target first and left that path uncovered.

Closes #73