Fix n_devices max_memory caps

[abhinav-bellapu]

Fixes #1496.

Summary

• Replace the invalid n_devices default max_memory={0: "auto", ...} with concrete CUDA byte budgets for devices 0..n_devices-1.
• Keep user-supplied max_memory values unchanged.
• Update Bridge type hints to reflect that HuggingFace accepts integer byte budgets as well as strings.
• Add a mocked-CUDA regression test so the failing resolver path is covered on CPU CI.

Why

Accelerate rejects "auto" as a max_memory value with ValueError: size auto is not in a valid format. n_devices should continue to resolve to device_map="balanced", but the generated caps need to be concrete values accepted by HF/Accelerate.

Validation

• make format
• make check-format
• uv run --python 3.11 mypy .
• uv run --python 3.11 pytest tests/unit/utilities/test_multi_gpu_unit.py -q
• uv run --python 3.11 pytest tests/acceptance/model_bridge/test_multi_gpu_bridge.py::TestResolveDeviceMap -q
• uv run --python 3.11 pytest tests/acceptance/model_bridge/test_multi_gpu_bridge.py -q
• make unit-test