add fp64 gradcheck for geglu activation and gate+up paths

[tejasexpress]

Base branch: please target dev, not main. GitHub defaults the base to
main; change it in the dropdown above if it is not already dev.

dev doesn't exist on the repo right now, only main, so this targets main.

What this changes

fixes #11 .. adds torch.autograd.gradcheck at fp64 for both geglu paths := ForgeGEGLUFunction (activation-only, separate gate/up) and ForgePackedGEGLUFunction (gate+up fused). runs through the Triton kernel, similar approach as RoPE's fp64 gradcheck. small shapes (2,3,8) and both approximate="tanh" and "none".

Numbers

n/a, no math changed

Correctness

• uv run ruff check .
• Relevant pytest suites pass
• Numerics checked against a reference, not just "it runs"

Anything a reviewer should know

using preserve_inputs=True in both new tests. With the default False,
gradcheck fails with "backward is not reentrant" becaus geglu_backward writes
the gradient into the same storage it saved for backward (dgate = gate
when preserve_inputs=False), so calling .backward() twice on the same
graph reads back corrupted values on the second call. that's the
default geglu() uses, so retain_graph=True + two backward calls would
silently give wrong gradients. didn't do anything abt it here, out of scope for
#11 , can open a separate issue if wanted.

tested on Colab Tesla T4