Re-enable FlashInfer for Llama4 on Blackwell in e2e fusion tests

[Copilot]

Purpose

Fix #28604 (resolved by #28739) - re-enable FlashInfer as the attention backend for Llama4 on Blackwell platforms in e2e fusion tests.

Test Plan

Existing e2e fusion tests in tests/compile/test_fusions_e2e.py will validate the change:

• test_attn_quant - Tests attention+quant fusion with FlashInfer on Blackwell
• test_tp2_attn_quant_allreduce_rmsnorm - Tests multi-GPU fusion patterns
• test_tp2_attn_quant_async_tp - Tests async TP with FlashInfer

Test Result

Changes are minimal and follow the existing pattern used for Llama3. The Llama4 model configuration now uses:

backend=AttentionBackendEnum.FLASHINFER if is_blackwell() else AttentionBackendEnum.TRITON_ATTN

This enables FlashInfer on Blackwell (where #28739 fixed the issue) while preserving TRITON_ATTN on Hopper (where #28568 remains unresolved).

Changes:

• Llama4 backend: conditional FlashInfer on Blackwell, TRITON_ATTN otherwise
• Removed TODO and [Bug]: Llama4 on B200 flashinfer produces garbage #28604 reference (fixed)
• Cleaned up stale Llama3 workaround comment
• Preserved [Bug]: FlashInfer attention backend on Hopper fails with llama4-scout and llama3 with fp8 kvcache #28568 TODO (Hopper kvcache=fp8 issue still open)

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

Original prompt

#28739 fixed an issue with llama4, can you please reenable flashinfer as the attention backend for llama4 on e2e fusion tests, and remove the issue links to #28604 which was fixed

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.