[cuda] Replace L1 access with warp shuffles in layer_norm_grad_input_kernel

[valentinandrei]

Replaces accessing L1 for gamma_val, c_h and c_loss with a warp shuffles. This is guaranteed to work as the unroll factor is lower than the warp width. On average it brings speedups between 0 and 25%.

We measured using the benchmark described in #107287

Fixes #ISSUE_NUMBER

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/110203

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

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

✅ You can merge normally! (6 Unrelated Failures)

As of commit 1e7947b with merge base 1e7947b ():

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

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

• Upload test stats / Upload test stats for 6477135208, attempt 1 (gh)
• Upload test stats / Upload test stats for 6477141481, attempt 1 (gh)
• Upload test stats / Upload test stats for 6477146426, attempt 1 (gh)
• Upload test stats / Upload test stats for 6477272978, attempt 1 (gh)
• Upload test stats / Upload test stats for 6477313221, attempt 1 (gh)
• Upload test stats / Upload test stats for 6477343396, attempt 1 (gh)

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