ParameterBuffer view replacement breaks Clone/serialization — needs scoped replacement

[ooples]

Bug Description

`NeuralNetworkBase.GetOrCreateParameterBuffer()` permanently replaces layer tensor fields (`_weights`, `_biases`, etc.) with buffer-backed view tensors via `SetTrainableParameters()`. This breaks:

1. Clone — GraphSAGE clone tries to reconstruct from the view tensors, which may have different metadata
2. Shape assumptions — MemoryNetwork's `TensorMatMul` gets rank-1 views instead of rank-2 tensors
3. Serialization — layers serialize the view tensor instead of the original data

Root Cause

`SetTrainableParameters(views)` replaces `_weights = views[0]` permanently. The replacement is meant for the tape to track operations, but it persists beyond the training step.

Proposed Fix

Scoped replacement — only replace during the training step scope:

```csharp
// Instead of permanent replacement:
trainable.SetTrainableParameters(views); // Mutates fields permanently

// Use scoped replacement that restores originals after training:
using var scope = trainable.UseParameterViews(views);
// ... tape forward + backward + update happens here ...
// scope.Dispose() restores original tensor references
```

This matches PyTorch where parameter buffers are contiguous storage but `nn.Parameter` objects maintain their identity.

Affected Models

MemoryNetwork (TensorMatMul rank mismatch), GraphSAGE (Clone constructor failure), DeepBelief (zero gradients), and potentially any model using Clone or serialization after training.

[ooples]

Audit Update (2026-04-06)

The ParameterBuffer view replacement pattern is still active in NeuralNetworkBase.TrainWithTape():

Line 2494: var initialParams = TapeTrainingStep<T>.CollectParameters(Layers, _parameterBuffer is null ? -1 : _layerStructureVersion);
Line 2495: var paramBuffer = GetOrCreateParameterBuffer(initialParams);
Line 2693: trainable.SetTrainableParameters(layerViews);

Current behavior: GetOrCreateParameterBuffer() creates a contiguous buffer and calls SetTrainableParameters(layerViews) on each layer, replacing their original tensor fields with buffer-backed views. These views share the underlying buffer memory.

Known broken paths:

1. Clone — layers try to reconstruct from view tensors which reference the shared buffer
2. Serialization — serializing a view tensor may serialize the entire buffer or produce incorrect results
3. Multi-threaded training — buffer is shared across all layers, concurrent access is unsafe

No scoped replacement mechanism exists yet. The buffer persists for the lifetime of the network (_parameterBuffer is a class field). There is no "restore original tensors" path.

Recommendation: Add a IDisposable scoped replacement that restores original tensors when the training step completes, or use a copy-on-write pattern for the buffer views.