Add interrupts, memory caps, async streaming, ETF callback results

[benoitc]

Closes four gaps that blocked running Python you do not control.

Interrupt running Python

py:interrupt/1 raises KeyboardInterrupt in the thread executing a context;
the in-flight call returns {error, interrupted}. py_context:call/eval/exec
interrupt automatically when their timeout expires, so {error, timeout} now
stops the Python code instead of abandoning the reply while the thread kept
burning CPU and the context stayed wedged.

The BEAM is the watchdog: no C-side timer thread, one new NIF. Works in both
worker and owngil mode (the latter attaches a fresh PyThreadState to the
context's subinterpreter), and is callable while the context process is blocked
in a NIF. Async exceptions land at bytecode boundaries, so code blocked in a C
call is interrupted once that call returns.

Per-context memory caps

py_context:new(#{mode => owngil, memory_limit => Bytes}). Accounting comes
from obmalloc arena traffic, which is the only place a free hook receives the
block size, so no per-object header is needed. Opt-in via
{enable_memory_limits, true}, since the allocator is hooked before Python
starts.

Enforcement raises MemoryError asynchronously rather than returning NULL from
the arena allocator: obmalloc treats a failed arena as a reason to fall back to
PyMem_RawMalloc, so a NULL would let the allocation silently succeed
off-arena and the cap would stop counting instead of stopping the code. Measured
during development: a 64 MB cap let usage reach 82 MB with no error.

Limits are documented: allocations over 512 bytes bypass obmalloc and are not
counted, granularity is one 1 MB arena, and worker mode returns
{error, memory_limit_requires_owngil} because those contexts share the main
interpreter.

Async generator streaming

py:stream_start/3,4 drives async generators on a private event loop. The docs
have claimed this since 3.0.0 with no code behind it; py:stream/4 with kwargs
and py:stream_eval/1,2 stay sync-only, now stated explicitly.

Callback results as external term format

Results returned from an Erlang callback now cross as term_to_binary and are
decoded by the same term_to_py converter used for call arguments, replacing a
Python repr string parsed with ast.literal_eval. Fixes binaries containing
backslashes, quotes, newlines or tabs (which produced an unparseable literal and
were silently delivered to Python as the raw repr text), [] arriving as '',
float precision loss, and the base64 round-trip for pids and refs, which now
cross as native Pid/Ref objects. Removes ~200 lines of duplicated encoders.

Breaking: an Erlang string returned from a callback ("abc", a list of
integers) now reaches Python as [97, 98, 99] rather than 'abc', matching how
call arguments have always converted. Return a binary for a Python str.

Performance

Callback round-trips get faster; hot paths are unchanged. The added per-request
cost is two uncontended mutex pairs, below the noise floor of a 9-15 us
operation.

	main	branch
ctx eval	14.9 us	16.4 us (within run-to-run variance)
ctx call	8.8 us	9.3 us (within run-to-run variance)
callback, nested map	104.3 us	45.8 us
callback, 8 KB binary	75.3 us	48.1 us