feat(entity): expose sensor and body tensor reads

[TATP-233]

Summary

• add the next NpEnv removal P1: tensor-native Manager-Based TorchEnv runtime #1703 Entity tensor facade for explicit named sensor reads
• add entity-ordered world-frame body tensor reads over the canonical backend track_* sensor contract
• validate execution capability, exact device, dtype, shape/layout, and finite values while preserving the public SimBackend boundary

Scope boundary

This is the narrow public Entity read API. Broad CPU host-bridge Manager execution must still aggregate these requests into one packed host-bridge plan; it must not call the per-name path from each term.

Validation

• uv run ruff format src/unilab/base/entity.py tests/envs/mdp/test_joint_position_action.py
• uv run ruff check src/unilab/base/entity.py tests/envs/mdp/test_joint_position_action.py
• uv run mypy src/unilab/base/entity.py
• make check
• uv run pytest -q tests/base/test_entity_facade.py tests/envs/mdp/test_joint_position_action.py tests/managers tests/envs tests/tasks — 505 passed, 8 skipped

Closes part of #1703

[TATP-233]

Validation at head 34cb90b5:

• local make check: pass (ruff/mypy/pyright; pyright has the existing optional drake import warning)
• focused suite uv run pytest -q tests/base/test_entity_facade.py tests/envs/mdp/test_joint_position_action.py tests/managers tests/envs tests/tasks: 505 passed, 8 skipped
• remote Focused CPU tensor profile: pass in 3m5s
• Gated single-GPU tensor profile: skipped as expected

Scope note: this adds the narrow public per-name Entity facade. Broad host-bridge Manager execution still needs the scene-level packed aggregation from these names and must not call this API per term.

[TATP-233]

Review-response update at 3eb0f3d0:

• normalize device once through torch.device(device); capability negotiation, backend reads, and exact tensor validation now use the same resolved device (fixes the str/unindexed CUDA mismatch risk);
• split dtype and exact-device validation diagnostics;
• cache resolved body layouts per requested-name tuple, using the existing strict _readonly_ids path (shape, integer dtype, negative, duplicate checks);
• fail closed on direct HOST_BRIDGE body tensor reads: CPU execution requires the next scene-owned packed host-bridge plan, so the facade no longer exposes a 4×B scattered-H2D path;
• fail closed on multi-name HOST_BRIDGE direct sensor reads for the same reason; single-name reads remain the narrow diagnostic path;
• tests now cover all four body fields, string-device input, host-bridge fail-closed behavior, and a device-resident fake profile.

Remaining accepted follow-ups for the next slice rather than this PR: public UniSim sensor-name capability metadata, entity-qualified body alias negotiation, tensor phase-scoped caching, and the actual scene-owned packed host-bridge plan.