perf: binding variable conversion (python and nodejs)

[stefan-gorules]

Faster Node & Python bindings: direct Variable to native conversion

Replaces the Variable -> serde_json::Value -> JsValue/PyObject round-trip with a direct, Rc-deduplicated conversion on every engine/decision eval-output path. Shared sub-trees (huge in traces, where node N's output is node N+1's input) are materialized once and reused.

Results (vs Rust floor; same machine)

fixture	node	python
expression-table-map +trace	18.6ms -> 3.1ms (6.0x)	11.1ms -> 2.1ms (5.4x)
dynamic-shipping +trace	29.2ms -> 4.4ms (6.6x)	16.9ms -> 2.6ms (6.5x)
customer-eligibility +trace	21.4ms -> 2.6ms (8.3x)	12.7ms -> 1.7ms (7.4x)
heavy graphs (no trace)	1.3-1.5x	1.7-1.8x

What's in here

• Direct converters for all eval paths: Node evaluate / safeEvaluate / evaluateBatch, Python sync + async evaluate (engine + decision).
• Same-thread (Python sync) = direct Variable -> PyObject; cross-thread (Node, Python async) = Send dedup arena built on the worker, materialized with handle reuse on the JS/GIL thread (Variable is !Send).
• New: Python evaluate_batch + async_evaluate_batch (parity with Node), parallel, returns [{success, data, error}].
• New: Python CI test job (ubuntu+macos x py3.9/3.13); release now gated on it.

[stefan-gorules]

Benchmark: before vs after (per-op, lower is better)

fixture	rust	node before -> after	python before -> after
decision-table-discounts (graph)	64.2 µs	451.2 -> 459.8 µs (0.98x)	106.3 -> 125.1 µs (0.85x)
expression-table-map (graph)	2,600.5 µs	41,226.7 -> 6,892.9 µs (5.98x)	29,395.1 -> 4,209.2 µs (6.98x)
credit-limit-adjustment (graph)	209.3 µs	858.2 -> 695.2 µs (1.23x)	357.4 -> 283.9 µs (1.26x)
dynamic-shipping-cost-calculator (graph)	3,160.2 µs	61,447.8 -> 9,816.9 µs (6.26x)	42,850.5 -> 5,210.6 µs (8.22x)
customer-eligibility-engine (graph)	2,270.9 µs	45,938.7 -> 5,824.2 µs (7.89x)	25,776.0 -> 3,612.3 µs (7.14x)
simple-assertion (policy)	43.5 µs	359.2 -> 385.4 µs (0.93x)	99.8 -> 114.3 µs (0.87x)
expression (policy)	73.2 µs	412.7 -> 444.4 µs (0.93x)	139.4 -> 147.3 µs (0.95x)
multi-assertion (policy)	743.5 µs	1,766.6 -> 1,811.3 µs (0.98x)	1,255.9 -> 1,118.3 µs (1.12x)
expression-chain (policy)	653.5 µs	2,664.0 -> 2,730.4 µs (0.98x)	2,064.8 -> 1,653.5 µs (1.25x)
match (policy)	64.6 µs	390.1 -> 400.5 µs (0.97x)	124.9 -> 136.1 µs (0.92x)

Big graph fixtures: ~6-8x faster (node and python). Small policy fixtures are flat (within run noise) - they're dispatch-bound, not marshaling-bound.