fallback: count buffer bytes independently of memoryview shape

[zhechen-cs]

The fallback sizes its unpackb limits with len(packed) and checks incoming feeds with len(view). For a multidimensional memoryview, that is the first dimension rather than the byte count. A six-byte message reshaped to (2, 3) fails to unpack with a spurious length error, while feeding it can exceed max_buffer_size without raising BufferFull. A zero-dimensional view raises TypeError on Python 3.12.

Normalize input to a contiguous, one-dimensional byte view before computing lengths. Contiguous input is cast without copying; non-contiguous input uses the byte copy already needed by feed. Decoding remains unchanged. Tests cover several shapes, non-contiguous multidimensional input, scalar views, exact-fit feeds, and preserving the stream after an oversized feed is rejected.

Validation on macOS arm64, CPython 3.12:

• All six new cases fail on the unmodified fallback and pass on the unmodified C extension.
• After the fix: fallback full suite 148 passed, 4 skipped; C extension full suite 151 passed, 1 skipped.
• PyPy 3.11: 138 passed, 5 skipped with test/test_except.py excluded. That unchanged module cannot be collected because this PyPy build does not provide _tracemalloc.
• Ruff lint and format checks, and git diff --check, pass.

The existing thread_unsafe marker warning also occurs on the base. Other platforms have not been tested locally.