Bazel build `acl` does not work for `torch-blade`

As we have found in [previous pr](https://github.com/alibaba/BladeDISC/pull/318).
Bazel build `acl` will cause torch-disc core dumped when calling function from `acl`. This wired core dump does not have a proper backtrace for us debug the details, we can only determine the problem comes from `libarm_compute.a` generated by bazel `make` rule.

`acl` have 3 output libraries, only `libarm_compute.a` diffs from the one generated by directly calling `scons`. However bazel `make` rule also calls scons with the same compilation options. The other two generated libs are the same by `md5sum` values.

Another wired thing is, this `bazel make rule` generated `acl` works well for `tensorflow-disc`.

We leave this issue here as a reminder for this bug. When this is fixed, we can ditch all the `cmake` calls from `common_setup.py` for `torch-blade`.

Start with the Bazel make rule, the generated libarm_compute.a, and common_setup.py, then compare the Bazel and direct scons outputs referenced in the issue and PR 318. Reproduce the torch-blade crash and verify why the Bazel-built ACL works for tensorflow-disc but not torch-disc; done means the ACL build no longer crashes torch-blade and the cmake calls can be removed.

Written by the indexing model from the issue text.