benchmark: add RTX 5060 Ti RMSNorm results

[alifurkanstahl]

Summary

Add 32 RMSNorm benchmark rows collected on an NVIDIA GeForce RTX 5060 Ti.

The data covers:

• Hugging Face and Liger providers;
• forward, backward, and full-pass latency;
• full-pass memory;
• BF16 with hidden size 4096;
• token-length sweep points 1024, 2048, 4096, and 8192.

Why

The shared benchmark dataset currently contains RMSNorm measurements from an A100 only. These rows add a consumer Blackwell data point without replacing the existing A100 history.

This data-only change is intentionally separate from the small-hidden RMSNorm kernel tuning work. The measurements use hidden size 4096 and should not be interpreted as validation of the block-row path used at hidden sizes up to 512.

Validation

• CSV parsed successfully with the repository schema.
• All 2,480 benchmark keys are unique.
• All 32 added measurements have finite median, 20th-percentile, and 80th-percentile values.
• The existing 48 A100 RMSNorm rows remain present.
• Diff contains exactly 32 additions and no deletions.
• git -c core.whitespace=cr-at-eol diff --check passed; cr-at-eol matches the CSV writer's existing CRLF output.

Hardware: NVIDIA GeForce RTX 5060 Ti. Measurements were produced with Liger Kernel 0.8.1 on 2026-08-06.