[python] Avoid recursive copies of flat CWD vectors

[YannByron]

Purpose

This PR remains draft: the long-window loader improvement is reproducible locally, but the target training-throughput improvement is not established.

CWD deep-copies each requested cell before applying transforms. Long action windows repeatedly recurse through the immutable scalars in numeric vectors. Copy flat lists of exact builtin immutable values directly; retain deepcopy for nested values, subclasses and custom objects so transforms and padding remain isolated.

This two-file follow-up to #10567 adds no sampler or cache configuration. The existing isolation test now checks overlapping/duplicate windows, repeated offsets, nested lists and mutable constant padding against explicit expected values.

Five alternating pairs on real RoboMIND JPEG data, after a discarded warmup round (samples/s; change is the ratio of medians):

Workload	Base 220f0fe	This change	Change
Loader, horizon 32, 0 workers	170.39	170.13	-0.15%
Loader, horizon 128, 0 workers	142.04	150.71	+6.11%
Loader, horizon 32, 2 workers	307.06	309.23	+0.71%
CPU ACT, horizon 32, 0 workers (20 steps)	21.89	22.05	+0.73%
CPU ACT, horizon 32, 2 workers (20 steps)	26.02	25.47	-2.13%
CPU ACT, horizon 128, 0 workers (20 steps)	20.75	20.93	+0.85%
CPU ACT, horizon 32, 2 workers (200 steps)	24.97	25.30	+1.33%

The 128-step loader's paired gains were +2.9% to +9.4% (median +6.4%). The 32-step loader has no stable gain. CPU training results do not establish a general training speedup or the target of exceeding HDF5/LeRobot by 30%. The longer 200-step check's paired gains were -2.73% to +3.67% (median -1.67%); all short-run results are retained above.

Measurement scope: M2 Pro / 32 GiB, Python 3.13, Torch 2.10, PyArrow 19, Python reader, local files and uncontrolled OS cache. Loader: four complete episodes / 2,299 frames, three original 640x480 JPEG views, unchanged sample order and float32 tensors, batch 16, four warmup and 64 measured batches. ACT: two training episodes, batch 2, a reduced CPU model with the same 64x80 resize for all readers; 20 measured steps after four warmup updates, plus the 200-step check after ten warmup updates. These are engineering measurements, not GPU or convergence validation.

Tests

• Real HDF5, Paimon and official LeRobot image readers produce identical sample IDs and tensors in an independent replay through the same worker/collate path. Every ACT run also agrees on normalization, per-step losses and initial/final model hashes within its configuration.
• CWD dataset/plan, window transforms, ACT benchmark, torch read and multimodal video regression scopes: 151 passed. One local native-plan test fails identically on unmodified master because the older installed native wheel lacks Plan.stats; current-wheel coverage is delegated to CI.
• The extended isolation test rejects both no-copy and nested shallow-copy mutations; the original deepcopy baseline passes.
• Focused flake8 and git show --check pass. Independent copy-semantics and benchmark-methodology review completed.
• CI Python 3.10-3.13 each fail only the same existing native-plan test: the released binding has no Plan.stats(). This is the master compatibility regression addressed by the still-open [python] Plan natively with pypaimon-rust releases that lack Plan.stats() #10620; that fix is not included here. CI is not yet green.