[AliAbouelazm]

Related to #2243. This fixes nested Sequence parsing, not the separate bare top-level Sequence(Scalar/Text) serialization error, so it does not close that issue.

Nested sequences use ragged flat values and row lengths even when their lengths are fixed. Select the matching TensorFlow parse specification and reconstruct NumPy sequences from stored row lengths, preserving trailing tensor dimensions and empty-row shapes.

Validation:

• Added 324 round trips covering nesting depths 2 and 3, variable/fixed/zero lengths, scalar/vector/matrix values, and NumPy/eager/graph parsing. Before the fix, 264 failed; all now pass.
• Final parser, serializer and sequence feature suites: 372 passed, 3 skipped, 3,140 subtests passed.
• Broader feature testing: 654 passed, 22 skipped, 43,878 subtests passed; 320 image-representation subtests failed due to Pillow 13 compatibility in unchanged code. Representative failures reproduce with the baseline parser. This broader run is not a clean pass.
• Whitespace checks pass. Pylint retains baseline diagnostics and outdated configuration warnings, so it is not a clean lint pass.
  Tested on Linux with Python 3.12 and TensorFlow CPU 2.21. Full repository CI was not run locally.