[None][perf] AutoDeploy: fuse FP8 quant into allreduce+residual+RMSNorm

[MrGeva]

Adds a new transform fuse_allreduce_residual_rmsnorm_quant_fp8 that runs after fuse_allreduce_residual_rmsnorm and folds the downstream per-tensor FP8 input quantize into the AllReduce kernel by switching from AllReduceFusionOp.RESIDUAL_RMS_NORM to
RESIDUAL_RMS_NORM_QUANT_FP8. This matches what the PyTorch backend does for FP8 dense models (see modeling_llama.py).

For each layer with TP > 1 we previously executed:
allreduce + residual + rmsnorm -> 1 fused kernel (bf16 out)
static_quantize_e4m3_per_tensor -> 1 extra kernel (fp8 out)
trtllm_fp8_prequant_linear -> matmul
With the transform the first two become a single AllReduce-fused
kernel emitting FP8 directly into the next linear's input.

Llama-3.1-8B-Instruct-FP8 on B200 TP=2, ISL=OSL=1000 (vs PT-backend):

• c=1: 2.42 -> 2.32 ms (-0.10) -- AD now ahead of PT (2.35)
• c=2: 2.42 -> 2.32 ms (-0.10) -- AD ahead of PT (2.37)
• c=16: 2.85 -> 2.74 ms (-0.11) -- half of the residual gap
• c=256: 24.10 -> 27.01 ms (+2.91) -- regression

Because the FP8-fused AllReduce kernel trades launch savings for less-optimal large-batch work, the transform is left enabled: false in default.yaml and opt-in per registry yaml. Accuracy on MMLU/GSM8K is unchanged from the non-fused FP8 path.

@coderabbitai summary

Description

Test Coverage

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.