Fix failing python tests on Windows

[zcbenz]

Some small fixes making all tests pass on Windows:

• Initialize ArrayDesc::data_size otherwise empty arrays which usually do not call array::set_data would get garbage data.
• Only use tensor ops in grouped mm when tf32 is enabled.
• Avoid accessing outputs[1] in SDPA when output_logsumexp is false.
• Do not try to get vjp for mask in test_sdpa_grad.

[awni]

Looks great, thanks!