CHANGELOG 0.10.0: clarify sm100f fwd/bwd limits and CUEQ_TORCH_COMPILE modes

[hsadasivan]

Summary

• Clarifies that the new sm100f (CC 10.0/10.3) improvement is specifically the forward kernel (hidden_dim ≤ 256), while the backward acceleration limit (hidden_dim ≤ 128) remains unchanged from before
• Expands the CUEQ_TORCH_COMPILE note to document that it applies to the forward pass only, and lists the supported integer modes: 1 → "default", 2 → "reduce-overhead", 3 → "max-autotune", 4 → "max-autotune-no-cudagraphs"

Test plan

• CHANGELOG-only change; no functional code modified

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.