[Python] enforce read bounds in Table

[ManoharPaturi]

Problem

The Python Table accessors trust offsets that come from the buffer itself when reading untrusted data:

• struct.unpack_from silently accepts negative offsets, reading relative to the end of the buffer, so Indirect/Offset with a negative computed target returns data from the wrong position with no error,
• String slices Bytes[start : start + length], and Python slicing clamps out-of-range ranges, so a malformed length silently returns a truncated string (or an empty one for huge unsigned lengths like 0xffffffff),
• Vector returns a start position beyond the buffer without complaint.

Change

Adds a _CheckOffset(buf, offset, size, what) guard:

• Offset validates the resolved vtable position (negative soffset no longer wraps),
• Indirect validates the computed target,
• String validates start + length against the buffer size before slicing,
• Vector validates the computed start (an empty vector may start exactly at the end).

Malformed buffers now raise IndexError with a clear message instead of silently producing wrong values. Mirrors the read-bounds enforcement direction the Swift port has adopted (#9277), applied to the Python port (companion to the TS-port change in #9280).

Test plan

• New tests/py_table_bounds_test.py: negative/out-of-range indirect and vtable targets, clamped and huge-length strings, vector start past the end, plus valid-target sanity cases — 10/10 pass
• On master, the same tests fail with wrong values instead of raising — confirming the silent misparse behavior
• tests/py_flexbuffers_test.py unchanged and passing (60/60) after the change
• (py_test.py requires flatc-generated code and runs in CI)