elementwise_util: s/common/compute/ almost everywhere and deprecate SAME_AS_COMPUTE

[swolchok]

As the title says, this is mostly a few related find-replaces, plus marking SupportedTensorDtypes::SAME_AS_COMPUTE deprecated.

[swolchok]

Stack from ghstack (oldest at bottom):

• RFC: Compute only in int32/long/float/double for portable ops to save size #9635
• Strip size_test binaries and report their sizes in build_size_test.sh #9633
• Deprecate non-internal elementwise_util APIs #9621
• -> elementwise_util: s/common/compute/ almost everywhere and deprecate SAME_AS_COMPUTE #9613

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9613

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (1 Unrelated Failure)

As of commit aab3e03 with merge base 811352d ():

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / test-llava-runner-linux / linux-job (gh) (trunk failure)
  test_generated_output

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[swolchok]

I'm not sure of the best way to handle SAME_AS_COMMON/SAME_AS_COMPUTE. The problems are as follows:

• SAME_AS_COMPUTE is likely not what anybody meant; the output type is a function of the input types not a function of whatever precision we happen to compute in. (hence this diff deprecating it)
• We labelled the incoming type CTYPE_COMMON, but it's really the compute dtype (hence this diff changing the label).
• We don't actually know the common dtype. We can't add it to the elementwise_util interface without a breaking change because elementwise_util isn't marked internal. I'm not sure how to solve this part (but I'm sending a PR right now to deprecate the non-internal elementwise_util interface on grounds that it's demonstrably not stable yet).

[swolchok]

we should land this if and only if @manuelcandales agrees it is improving the code, since the diffs I was trying to support with it have been abandoned.