Fix madspace_tests failure on cuda: copy read-only arrays in test_beam_rapidity

[Copilot]

The manneback / madspace_tests job on the cuda runner fails in test_beam_rapidity (4 parametrizations). lab.map_inverse(...) raises TypeError: __dlpack__() got an unexpected keyword argument 'max_version'. I couldn't run the tests locally, so this fix is unverified.

Cause

• The test builds x1 and x2 with np.broadcast_to, which returns read-only, zero-stride arrays.
• The runner's numpy appears to be old. It can't export read-only arrays over DLPack without max_version, and it doesn't accept max_version either, so the retry in dlpack_to_tensor (function_runtime.cpp) fails with the TypeError.
• I inferred the numpy version from the error and haven't confirmed it on the runner.

Change

• madspace/tests/test_beam_frame.py: .copy() the broadcast x1 and x2 so they are writable and contiguous before they go to map_inverse.

Not changed

dlpack_to_tensor shares one dlpack_version_cache across all arguments of a call. In the fallback branch, a failure caused by an argument other than max_version (here, the read-only export error) is retried with max_version. That retry then raises the TypeError above, which hides the original error. Users passing read-only numpy arrays on an old numpy would hit the same failure. I haven't changed this.