Make spatiotemporal autoencoders reconstruct sequences

[sergioald]

Summary

This PR changes the two spatiotemporal autoencoders to reconstruct complete
input sequences.

The public mappings become:

ConvLSTMAutoencoder:
(B, T, C, H, W) -> (B, k) -> (B, T, C, H, W)

HybridConvLSTMTransformerAutoencoder:
(B, T, C, H, W) -> (B, k) -> (B, T, C, H, W)

There is no reconstruction_mode option. Both classes now have one clear
autoencoder objective: reconstruct the complete sample supplied to the
encoder.

Rationale

The previous implementation encoded a complete sequence but reconstructed
only its final input frame. That sequence-to-frame behaviour was useful, but
it was surprising for classes named autoencoders and prevented a direct
full-window comparison with PCA and the existing LSTMAutoencoder.

A new user would normally expect:

X_hat = model.predict(X)
assert X_hat.shape == X.shape

This PR makes that expectation true for all reconstruction-oriented sequence
models.

Architecture

ConvLSTM autoencoder

• Encodes the complete sequence with the existing ConvLSTM stack.
• Maps the final recurrent representation to one latent vector of size k.
• Adds learned temporal embeddings to the latent vector.
• Uses an LSTM temporal decoder to create one code per output timestep.
• Applies one shared spatial decoder to every timestep.

Hybrid ConvLSTM-Transformer autoencoder

• Preserves the existing ConvLSTM and temporal-attention encoder.
• Maps the full input window to one latent vector of size k.
• Uses latent-conditioned learned temporal queries.
• Refines the queries with the configured linear or standard attention blocks.
• Applies one shared spatial decoder to every timestep.

API behaviour

model = ConvLSTMAutoencoder(k=32)
model.fit(X)

Z = model.encode(X)
X_hat = model.predict(X)
X_from_z = model.decode(Z)

assert Z.shape == (len(X), 32)
assert X_hat.shape == X.shape
assert X_from_z.shape == X.shape

The same behaviour applies to
HybridConvLSTMTransformerAutoencoder.

fit(X) and the reconstruction metric helpers use the complete sequence X
as their default target. A custom y must have the same shape as X.

Breaking change

This intentionally changes the output shape of both spatiotemporal classes:

before: (B, C, H, W)
after:  (B, T, C, H, W)

Existing final-frame checkpoints are not compatible with the new decoder
architecture. They should remain tied to the earlier BlueMath_tk version that
created them. Loading them into the new classes is expected to fail rather
than silently initialise missing sequence-decoder parameters.

The removed final-frame operation is better represented in future by a
separately named sequence-to-frame model, not by an autoencoder mode.

Tests

The PR adds or updates coverage for:

• complete-sequence output shapes for ConvLSTM and Hybrid models;
• predict(X) == decode(encode(X));
• automatic full-sequence reconstruction targets;
• manual reconstruction-error agreement;
• frame-target rejection;
• fixed sequence-length validation;
• sequence checkpoint round trips;
• one-vector decoding;
• non-divisible spatial dimensions and output cropping;
• linear and standard Hybrid attention;
• finite non-zero gradients in sequence-specific decoder parameters;
• timestep-dependent decoder outputs;
• rejection of the removed reconstruction_mode argument;
• updated public API and smoke-test expectations.

Local validation

Replace these placeholders with the results from the complete checkout:

compileall: passed
focused Ruff checks: passed
sequence reconstruction tests: 16 passed
autoencoder API tests: 20 passed
autoencoder smoke tests: 8 passed
metrics tests: 8 passed
full suite: 257 passed, 113 warnings