📰 ciflow/torchtitan/186752: [Inductor] Host-side TMA descriptors for Blackwell mm templates

Add a {{tma_descriptor()}} Jinja hook that lets mm templates declare TMA descriptors declaratively. When config.triton.enable_host_side_tma is True, the hook registers the input's pointer arg in host_tma_descriptor_args so the launcher (from PR #185825) replaces it with a TensorDescriptor at launch time. When False, the hook emits device-side...