[FEATURE SUPPORT] add geru python kernel

[LoserCheems]

Summary

• Purpose: Add a Python reference implementation of the GERU (general rank‑1 update) kernel so other backends and tests have a clear numerical baseline for A = A + alpha * x yᵀ.
• Issue: Implements the work tracked in [FEATURE REQUEST] geru Python kernel implementation #25.

Design

• API: geru(A: torch.Tensor, x: torch.Tensor, y: torch.Tensor, alpha: float) -> torch.Tensor (matches other Python ops).
• Math: Computes the outer product and updates A via A = A + alpha * x[:, None] * y[None, :].
• Semantics: Returns an updated tensor (out‑of‑place result) to avoid in‑place aliasing surprises.

Changes

• Added: geru.py — Python helper implementing the rank‑1 update with a clear docstring.
• Added: geru.md — documents math, shared API, and points to implementations/tests.

Implementation notes

• Device/dtype: Accumulates on the same device and dtype as inputs (no host transfers).
• Simplicity: Uses broadcasting (x[:, None] * y[None, :]) for clarity and correctness rather than micro‑optimizations.
• Validation: Relies on PyTorch for shape/dtype errors to keep the reference minimal.

Tests

• Status: No unit tests in this PR.
• Manual check: Verified against small CPU examples and numpy/PyTorch equivalents.
• Next: Add tests/test_geru.py to use the Python helper as ground truth and validate future PyTorch/Triton/CuTe backends.

Documentation

• Docstring: Present in geru.py.
• Guide: geru.md includes the definition, interface, and where to find implementations/tests.

Checklist

• Linked issue provided ([FEATURE REQUEST] geru Python kernel implementation #25)
• API stabilized
• Tests added or updated (planned)
• Docs added or updated (geru.md)
• No known performance regressions

Would you like me to add the basic tests/test_geru.py now (correctness + device/dtype parametrization)?