Build CUDA plugin before torch and torch_xla in nightly build

[will-cromar]

The nightly CUDA package build is failing with this error:

Step #2 - "build_xla_docker_image":       Error in repository_rule: invalid repository name '@pypi_torch_nightly+20240722_cp311_cp311_linux_x86_64_whl'

AFAICT this is happening because the build_srcs step is producing incorrectly-named wheels (see #7697). The GPU plugin build has zero dependency on PyTorch at all -- it builds completely separately in our CI. I believe it is unfortunately getting pulled in transitively through OpenXLAs build rules:

xla/WORKSPACE

Line 100 in 5ea4535

local_wheel_workspaces = ["@torch//:WORKSPACE"],

Move the plugin build to happen first as a temporary fix since there's no dependency between these steps.

Triggering the nightly build manually to see if this works

[will-cromar]

The test build passed. This PR should unbreak the GPU build.