Add process-bound Python environments

Each Erlang process gets its own Python globals/locals dict stored as a
NIF resource in the process dictionary. When the process exits, the
resource destructor frees the Python dicts.

Works in both worker and subinterpreter modes.

- Add py_env_resource_t and create_local_env NIF
- Add _with_env variants of context_exec/eval/call
- Add get_local_env/0 to py.erl using process dictionary
- Update py_context to pass EnvRef to NIFs
- Add tests for isolation, __main__ access, persistence, cleanup

Author: benoitc

benchmark_results/reactor_modes_20260309_093558.txt

@@ -0,0 +1,76 @@
+
+========================================================
+  Reactor Modes vs Channel API Benchmark
+========================================================
+
+System Information
+------------------
+  Erlang/OTP:       28
+  Schedulers:       14
+  Python:           3.14.3 (main, Feb  7 2026, 05:34:13) [Clang 17.0.0 (clang-1700.6.3.2)]
+  Subinterp:        true
+
+
+--- Channel API ---
+  A) Messages on persistent channel (1000 msgs):
+    Size     64:      0.2 us/op,  5988024 ops/sec
+    Size   1024:      0.3 us/op,  3508772 ops/sec
+    Size  16384:      0.9 us/op,  1169591 ops/sec
+
+  B) Full lifecycle per op (create+send+recv+close, 200 ops):
+    Size     64:      0.3 us/op,  3174603 ops/sec
+    Size   1024:      0.4 us/op,  2631579 ops/sec
+    Size  16384:      1.1 us/op,   877193 ops/sec
+
+--- Reactor (Worker Mode) ---
+  A) Messages on persistent connection (500 msgs):
+    Size     64:     66.2 us/op,    15110 ops/sec
+    Size   1024:     66.8 us/op,    14972 ops/sec
+    Size  16384:     73.1 us/op,    13688 ops/sec
+
+  B) Full lifecycle per op (connect+handoff+echo+close, 100 ops):
+    Size     64:    232.3 us/op,     4305 ops/sec
+    Size   1024:    230.3 us/op,     4343 ops/sec
+    Size  16384:    240.2 us/op,     4164 ops/sec
+
+--- Reactor (Subinterpreter Mode) ---
+  A) Messages on persistent connection (500 msgs):
+    Size     64:     65.9 us/op,    15172 ops/sec
+    Size   1024:     66.9 us/op,    14947 ops/sec
+    Size  16384:     73.7 us/op,    13573 ops/sec
+
+  B) Full lifecycle per op (connect+handoff+echo+close, 100 ops):
+    Size     64:    216.6 us/op,     4618 ops/sec
+    Size   1024:    247.0 us/op,     4049 ops/sec
+    Size  16384:    213.0 us/op,     4695 ops/sec
+
+========================================================
+  COMPARISON SUMMARY
+========================================================
+
+A) PERSISTENT CONNECTION (messages on existing connection)
+-----------------------------------------------------------
+    Size |      Channel |    Reactor/W |    Reactor/S
+----------------------------------------------------
+      64 |      5988024 |        15110 |        15172
+    1024 |      3508772 |        14972 |        14947
+   16384 |      1169591 |        13688 |        13573
+
+B) FULL LIFECYCLE (create + send/recv + close per op)
+-----------------------------------------------------------
+    Size |      Channel |    Reactor/W |    Reactor/S
+----------------------------------------------------
+      64 |      3174603 |         4305 |         4618
+    1024 |      2631579 |         4343 |         4049
+   16384 |       877193 |         4164 |         4695
+
+Legend:
+  Channel   = py_channel API
+  Reactor/W = erlang.reactor with worker mode
+  Reactor/S = erlang.reactor with subinterpreter (SHARED_GIL)
+
+Notes:
+  - A) measures throughput on persistent connection (best case)
+  - B) measures full lifecycle including setup/teardown
+  - Channel is queue-based, Reactor goes through TCP stack
+