fix(numeric): support legacy PyTorch autocast fallback

[PinkTulips139]

Summary

Adds a backward-compatible autocast-disabled context for legacy PyTorch versions that do not provide the top-level torch.autocast API.

The issue was observed in the repository's Python 3.8 / PyTorch 1.8 CI environment, where existing Mixture, MoA, and MoT tests failed with AttributeError: module 'torch' has no attribute 'autocast'.

Changes

• Preserve the existing torch.autocast(device_type=..., enabled=False) path on modern PyTorch.
• Fall back to torch.cuda.amp.autocast(enabled=False) for legacy CUDA environments.
• Use nullcontext() for legacy CPU, MPS, and unsupported device types.
• Keep compatibility handling centralized in disabled_autocast() rather than patching individual Mixture modules.
• Add regression tests for modern CPU/CUDA/MPS paths, legacy CUDA, legacy CPU/MPS, and unknown devices.

Scope

This PR does not change:

• routing behavior;
• model outputs;
• optimizer logic;
• training configuration;
• LoRA behavior;
• CI workflow definitions.

It only provides a compatibility fallback for the shared numerical helper.

Validation

• 8 passed in tests/test_torch_legacy_compat.py
• 13 passed in the previously affected Mixture test files
• Ruff passed
• Python compilation passed
• git diff --check passed

No GPU is required for the regression tests; the legacy CUDA path is validated with mocks.