Fix Sequential model serialization format

[sdatkinson]

Summary

• serialize Sequential as a true container whose config.models entries are complete child NAM objects
• keep the wrapper's top-level weights empty so each child owns its architecture, configuration, weights, and sample rate
• add recursive import support for canonical and nested Sequential models
• require the case-sensitive architecture name "Sequential"
• reject non-empty wrapper weights and the unrecoverable legacy bare-child-config representation
• document the intentional compatibility break without changing the .nam file version

This aligns trainer exports with the Sequential support merged in NeuralAmpModelerCore#323 and supersedes the experimental serialization introduced in #596.

The branch is based on the current main, including the Linear tap-order fix from #697. Its chronological-tap test now verifies the weights inside each exported child rather than the removed flattened wrapper weights.

Tests

• python -m pytest -q tests/test_nam/test_models/test_sequential.py tests/test_nam/test_models/test_from_nam.py tests/test_nam/test_models/test_factory.py — 48 passed, 1 skipped
• python -m pytest -q --ignore=tests/test_graceful_shutdown.py --ignore=tests/test_bin — 382 passed, 21 skipped
• python -m black --check nam/models/_from_nam.py nam/models/sequential.py tests/test_nam/test_models/test_sequential.py
• python -m sphinx -W -b html docs/source build/docs-html

Compatibility note

Earlier experimental Sequential files stored bare child configs plus flattened top-level weights. Those files cannot be reconstructed reliably because they omit child architectures and weight boundaries, so this PR intentionally rejects them while retaining file version 0.7.0.